Note: Descriptions are shown in the official language in which they were submitted.
2~
COMPUTBR Y~TEN MEMORY PERFORMANCE
Ih~K~ sMENT APPARaTU~
FIELD OF THE lNv~ ON
This invention relates to computer memory systems
and, in particular, to computer system memory
performance improvement apparatus that both monitors
the data retrievability efficiency of the computer
system memory and modifies data set storage locations
in response to memory performance conflicts detected
in the computer system memory.
PROBLEM
It is a problem in data processing systems to
efficiently manage the storage of data sets on the
computer system memory. Many large computer systems
have a hierarchy of data storage devices used therein.
These data storage devices range from off line
magnetic tape cartridge data storage systems to direct
access storage devices which are directly connected
to the computer system and available on line and,
finally, cache memory which is a fast on line memory
for use with data sets that are frequently read by the
computer system. The difficulty with this hierarchal
arrangement of data storage devices is to dynamically
allocate the location of the data sets into these
various data storage devices so that the frequency o~
~0(~ 6
-
data retrieval from these data sets matches the
location and time wise retrieval efficiency of the
data storage device. It is advantageous for computer
system performance purposes to place the most
frequently retrieved data sets in cache memory and the
most infrequently retrieved data sets in the off line
magnetic tape cartridge data storage systems.
From surveys done in the data processing field,
in 1971 the median data processing system installation
had 1.2 gigabytes of direct access data storage device
capacity and these data storage devices had 450 data
sets. In 1984, surveys revealed that the typical or
median installation had 92 gigabytes of direct access
data storage device capacity which devices had 44,000
data sets stored therein. There is a continuing
significant growth in the capacity of direct access
data storage devices in the typical computer system.
However, the utilization of these direct access data
storage devices has fallen from 80% in the late 1960s
to approximately 55% in the present time frame. Thus,
there are significant increases in the cost of direct
access data storage device memory with a decreasing
efficiency in the use of these direct access data
storage devices in the hierarchy of data storage
devices in the computer memory.
It is a commonly used metric in the computer
system field that in order to manage the computer
system memory, it requires approximately 1 data
management employee to manage every 10 gigabytes of
data in a computer system installation. This data
management person performs the functions of I/0 load
balancing, data set placement, and cache tuning in
order to provide improved efficiency in the use of the`
direct access data storage devices in the computer
Z~06836
-
system. Any improvement in the management of the data
storage devices therefore reduces the number of data
management employees required to manage the computer
system memory.
Z~og8~6
80LUTION
These problems are solved and a technical advance
achieved in the art by the computer system memory
performance improvement apparatus. This apparatus
makes use of a knowledge based (expert) system to
monitor the performance of the computer system data
storage devices, identify memory performance conflicts
and resolve these conflicts by directing the
relocation of data sets to other segments of the
computer system memory.
The subject apparatus includes a number of data
bases which are used by the expert system software to
manage the computer system data storage devices. One
element provided in this apparatus is a set of data
storage device configuration data that provides a
description of the various data storage devices and
their interconnection in the computer system. A
second element is a knowledge data base that includes
a set of functional rules that describe the data
storage device management function. These rules
indicate the operational characteristics of the
various data storage devices and the steps that need
to be taken to provide the various improvement
functions required of the computer system memory.
Operating on these two data bases is an expert
system, which is a computer program that uses
explicitly represented knowledge and computational
inference techniques to achieve a level of performance
compatible to that of a human expert in that
application area or domain. The expert system
includes an inference engine that executes the rules
which comprise the basic intelligence of the expert
system. The inference engine allows a virtually
infinite number of rules or conditional statements to
20 06836
be chained together in a variety of ways. The
inference engine manages the rules and the flow of
data through those rules. Thus, the expert system
uses the data stored in the configuration data base
and the knowledge data base as well as data from
monitoring of the actual performance of the data
storage devices in the computer system to analyze on
a dynamic basis the performance of the computer system
data storage devices.
The expert system identifies memory performance
conflicts such as a performance degradation of the
computer system data storage devices due to a
plurality of computers in the computer system
attempting to access a common data storage device.
The expert system identifies the performance conflict
as well as the data sets stored on these data storage
devices related to this conflict. Once the data sets
involved in the performance conflict are identified,
the expert system determines alternative memory
storage locations for these data sets and activates
various software routines to transport these conflict
data sets to the alternative data storage locations.
The relocation of these conflict data sets resolves
the memory performance conflict and improves the
retrievability of the data stored on these data
storage devices. By performing the conflict
identification and resolution on a dynamic real time
basis, the data storage devices of the computer system
are operated in a more efficient manner and the
retrievability of the data stored on these data
storage devices is significantly improved without the
need for the data management personnel. The computer
system memory performance improvement apparatus
therefore, continuously monitors and modifies the
Z006~6
with the computer system.
~006836
BRIEF DE8CRIPTION OF THE DRAWING
Figure 1 illustrates, in block diagram form, the
architecture of the computer system memory performance
improvement apparatus;
Figure 2 illustrates in block diagram form, the
architecture of the cache tuning and DASD performance
optimizer elements of the computer system memory
performance improvement apparatus;
Figures 3 and 4 illustrate the elements that
comprise the cache tuning and DASD performance
optimizer elements of the computer system memory
performance improvement apparatus.
2Q~68:~6
-
DET~TT~D DE8CRIPTION
The computer system memory performance
improvement apparatus functions to monitor the
performance of a computer system memory and manage
S this computer system memory in order to maximize
performance. This function is accomplished by the
use of a knowledge based (expert) system that monitors
the activity of the various data storage devices in
the computer system memory, identify memory
performance conflicts and take steps to resolve these
conflicts in order to improve the performance of the
computer system memory. This apparatus includes a
number of databases which are used by the expert
system software to manage the data storage devices.
Data Processing 8ystem Architecture
Figure 1 illustrates a typical data processing
system in block diagram form, which system includes
the computer system memory performance improvement
apparatus. This computer system consists of a
mainframe 101 which can be any of a number of large
computer system such as an IBM 3084, an Amdahl 5860
or a NAS 9000 series computer. Each of these
mainframe computers can process on the order of 100,00
instructions per second and are typically equipped
with a significant amount of disk storage memory,
typically on the order of 30 to 100 gigabytes of data
storage capacity. For the purpose of illustration,
assume that mainframe 101 is an IBM 3084 type of
computer running an operating system 102 that
typically is the MVS/XA operating system. These
products are well known in the field and need no
further explanation herein. Mainframe 101 is
interconnected via a plurality of channel processors
Z()1)683~,~
` -
110 - 119 to the data storage devices associated with
mainframe 101. For the purpose of illustration, a
number of disk storage devices 150-0 to 150-15 are
illustrated in Figure 1. Each of these disk drives
are a commercially available device such as the 8380E
disk drive system or any other ones of the multitude
of similar devices. In a typical configuration,
sixteen disk drives 150-0 to 150-15 are connected via
associated data channels 140-0, 140-15 to a device
controller 130 which is typically an 8890 type of
controller. Device controller 130 functions to
multiplex the data being read to and from the
plurality of disk drives 150-0 to 150-15 onto a
channel path 120 which interconnects device controller
130 with channel processor 110.
The computer system can be equipped with a
plurality of channel processors illustrated in Figure
1 as devices 110 to 119. Each channel processor 110
is interconnected with mainframe 101 and provides
access to the disk storage devices 150-0 to 150-15 to
mainframe 101. While this block diagram illustrates
a general computer system configuration, it in no way
should be construed as limiting the applicability of
Applicants' invention. In particular, various memory
configurations can be used to interconnect mainframe
101 with various data storage devices. The disk drive
devices illustrated in Figure 1 are exemplary and many
computer systems rely on tape drives to implement data
storage devices. Included in the computer system
illustrated in Figure 1 is a cache memory 160 which
is part of device controller 130. Memory performance
improvement apparatus 103 is included in mainframe 101
as is operator terminal 170 which is connected via
data link 160 to mainframe 101.
200~836
._
Nemory Confi~uration and Manaqement
From a data management standpoint, the computer
memory illustrated in Figure 1 is viewed as a
subsystem 180 that consists of sixteen volumes of disk
storage 150-0 to 150-15. Each volume of this
subsystem contains a plurality of data sets or data
files. These data sets or data files are stored on
a corresponding volume in an ordered fashion by
dividing the volume up into sub-elements. For
example, in the 8380E disk drive, each volume consists
of 1770 cylinders each of which includes fifteen data
tracks. Each data track can store 47K bytes of data.
In operation, mainframe 101 accesses data stored
on the data volumes in subsystem 180 by transmitting
a data retrieval request via the channel processor 110
associated with the data volume example 150-0 that
contains the requested data. This data retrieval
request is transmitted by channel processor 110 over
channel path 120 to device controller 130, which in
turn manages the retrieval of the data from the
designated data storage device 150-0. This data
retrieval is implemented by device controller 130
monitoring the rotation of disk drive 150-0 to
identify the beginning of the data set stored on a
track of the disk storage system. Once the disk in
the disk storage system has rotated sufficiently to
place the beginning of the requested data set under
the read/write heads of the disk storage system, the
data is transferred from disk drive 150-0 over data
link 140-0 to device controller 130. Device
controller 130 temporarily stores the retrieved data
for transmission to mainframe 101 via channel path
120 and channel processor 110.
2006836
_.
11
It is obvious from the arrangement that this is
an asynchronous data transfer in that mainframe 101
transmits data retrieval requests to various device
controllers where these requests are processed as the
data sets become available. If too many data
retrieval requests are made via one route, such as
concentrating the requests on a specific volume or
small subset of volumes, the performance of the
computer system memory significantly degrades. It is
advantageous to equalize the data retrieval requests
throughout the various channel processors, channel
paths and disk drives. Thus, a distribution of the
data retrieval requests among the various volumes on
each subsystem and across all of the subsystems
significantly improves the memory performance. There
are many choices available to improve memory
performance such as moving data sets from one volume
to another or moving a volume from one subsystem to
another. It should also be noted that for purposes
of this patent disclosure, the terms "string" and
"subsystem" can be used interchangeably.
Another such improvement is the use of a cache
memory such as cache 160 which enables the more
frequently accessed volumes to load data in the cache
for more rapid data retrieval times. The volumes are
selected as being suitable for caching based on the
data sets that are stored therein. Any input and
output accesses to the data that use the cache memory
benefit from the cache by speeding up the data
transfer time. In order to preserve data integrity,
any writes that are performed cause the data to be
written directly to the disk and therefore these
operations gain no benefit from a cache memory. A
good volume for caching is one that has a high read
2(~0~ 6
-
12
rate, high activity and input/output accesses
concentrated in a small number of tracks rather than
scattered across the entire disk. Volumes that
satisfy these conditions have a high hit ratio and
therefore the input/output reads often find the
desired track in cache and do not have to spend the
time retrieving the data set from the disk as
described above. Often, subsystems contain no volumes
that are cachable or the activity of volumes that are
cachable accounts for only a fraction of the total
input/output activity of the subsystem.
Cache Performance Improvement
In order to gain full benefit of the cache, it
is possible to reorganize the subsystem in such a way
as to increase the amount of input/output accesses
that go to specific cache volumes. A significant
problem is to reorganize the subsystem on a data set
level so as to separate the possible caching data sets
from data sets that would degrade the cache
performance and, in the process, create entire volumes
that are suitable for storage in the cache memory.
It is obvious that restructuring the data set
organization can entail a complete restructuring of
the volumes of the subsystem which involves an
enormous number of data set moves to accomplish.
Instead of a complete reorganization of the subsystem,
a better approach is to identify places on the
subsystem where conflicts exist between possible cache
data and data that is detrimental to the cache. The
conflict exists where there is good cache data and bad
cache data on the same volume: if the volume where
cached in order to obtain the benefit of caching the
good data, the bad data would cause so much
Z0~6f~
-
13
interference in the cache that the overall performance
would actually deteriorate. The computer system
memory performance improvement apparatus locates
conflicts and then moves the smallest number of data
sets in order to resolve the conflict. Using this
approach, a small number of moves provides a large
increase in performance and better utilization of the
cache memory.
In order to determine which data sets are
possible candidates for caching and which data sets
should not be cached, the input/output activity of the
subsystem is monitored over a period of time and data
is recorded on a data set basis. This stored data
consists of the input/output activity (the number of
lS input and output operation to the data set), the read
percentage (percentage of operations that were reads),
and the locality of reference measure, which is
defined as the number of input and output to the data
set divided by the number of unique tracks covered by
those inputs and outputs over the period monitored.
Along with these numbers, is stored the name of the
data set, its size in bytes and a value called the
reason code that flags data sets that are not suitable
for moving. This data is used by the computer system
memory performance improvement apparatus in
identifying the data sets that are to be moved and the
target location for these data sets.
MemorY Performance Improvement Apparatus
The computer system memory performance
improvement apparatus consists of a set of integrated
data storage management functions that use knowledge
or expert system techniques to monitor and
automatically assist in the control of performance of
;~00Ç~ 6
the direct access data storage sub-systems. Figure
2 illustrates the principal components of the computer
system memory performance improvement apparatus 103.
Memory performance improvement apparatus 103
interfaces with and runs on the MVS operating system
102 of mainframe 101. The MVS operating system 102
provides device management, file management, test
management, processor management, communication
management and a number of other administrative
functions to the computer system memory performance
improvement apparatus 103.
In order to clarify terminology, the various
aspects of knowledge systems are discussed. Knowledge
systems are computer based systems that emulate human
reasoning by using an inference engine to interpret
the encoded knowledge of human experts that is stored
in a knowledge base. If the domain of the knowledge
base is sufficiently narrow and a sufficiently large
body of knowledge is properly coded in the knowledge
base, then the knowledge system can achieve
performance that matches or exceeds the ability of a
human expert. In such a case, the knowledge system
becomes an expert system.
One step in building an expert system is
obtaining and encoding the collective knowledge of
human experts into the machine readable expert system
language. The specific implementation details of this
encoding step is largely a function of the particular
syntax of the expert system programming language
selected. As an example one such expert system
programming language is PROLOG which is described in
the text "Programming In Prolog" by W. F. Clocksin and
C. S. Mellish, Springer Verlag Inc., New York, New
York (1981). Basically, the expert system contains
2(~0Çi8~
`_
a set of rules and instructions on how the rules are
to be applied to the available facts to solve a
specific problem.
In the Computer System Memory Performance
Improvement Apparatus 103, expert system techniques
are used to monitor the performance of the computer
system memory. The facts gathered through this
monitoring operation are then used to identify
modifications to the organization of the computer
system memory as well as the data stored therein to
improve the performance of the computer system memory.
Figure 2 illustrates some of the various routines or
subsystems that are implemented in computer system
memory performance improvement apparatus 103.
Included are cache volume creation function 105,
and DASD performance optimizer function 106. Cache
subsystems often end up in a state where there are no
volumes on the DASD subsystem suitable for caching,
or where the number of I/Os to cached volumes account
for only a fraction of the total number of I/Os to the
subsystem. The cache volume creation 105 creates
volumes (205) on the subsystem that are suitable for
caching. The cache volume creation 105 monitors (201)
the performance of the cache memory 160-169 and
identifies where cache candidate data sets are in
conflict with data sets that are unsuitable for
caching (202). Cache volume creation 105 analyzes the
data set conflicts and recommends data set movements
(203) that separate the good candidates from the bad
candidates thereby concentrating the cache candidate
data sets on a few cachable volumes. The cache volume
creation 105 analyzes one DASD subsystem each time it
is invoked by scheduler 109 and attempts to achieve
the maximum effect with the minimum number of data set
20~)6~3~
16
moves (204). DASD performance optimizer 106 makes
data set movement recommendations to improve the
performance and utilization of the DASD devices 150-
0 to 150-15. DASD performance optimizer 106 operates
in a proactive manner, moving data sets from over-
utilized volumes and subsystems to under-utilized
volumes and subsystems. DASD performance optimizer
106 monitors (211) the activity on the DASD devices
150-0 to 150-15 and identifies (212) potential DASD
performance problems caused by over-and under-utilized
volumes and subsystems. DASD performance optimizer
106 then relocates (213) data sets between volumes to
avoid performance problems.
Scheduler 109 controls the operation of all
functions and determines when activities can take
place. Parameters used to establish a schedule are:
installation size, peak processing times, list of
functions to execute, user provided constraints. Data
Collection 171 records the I/O activity of all devices
on a prescheduled basis (e.g. hourly totals). In
addition, detailed volume activity data and track data
can be recorded. Data collection 171 creates the
statistical information required by the remaining
modules of the system. Database 108 contains the
system configuration information as well as the
statistical activity data. Those skilled in this art
will appreciate that in the use of the term
"database", it is contemplated that both "knowledge"
and "configuration" are contained in the general data
repository. Database 108 is illustrated as being
stored on disk drive 150-15, although it could also
be stored in computer memory performance improvement
apparatus 103 on mainframe computer 101.
Z006~6
._
17
Cache Volume Creation
The Cache Volume Creation (CVC) module 105 is one
of the performance tuning routines of the computer
system memory performance improvement apparatus 103.
The cache volume creation module 105 uses Expert
System techniques to analyze cache subsystems and make
recommendations that improve input/output I/O
performance in the computer system memory.
Because of the difficulty involved with using
cache properly, or because of a lack of data
management employees to do the job, cache subsystems
often end up in a state where there are no volumes on
the subsystem suitable for caching, or where the
number of I/Os to cached volumes account for only a
fraction of the total number of I/Os to the subsystem.
The purpose of the cache volume creation module is to
remedy this situation by attempting to create volumes
on the subsystem which are suitable for caching.
There may very well be data sets on the subsystem
which would benefit from cache, but because the cache
operates on a volume basis and the data sets reside
on volumes which would be detrimental to I/O
performance if cached, these data sets can't be
cached. The job of the cache volume creation module
is to identify where cache candidates are in conflict
with data sets that are unsuitable for caching and
recommend movements that separate the good candidates
from the bad candidates, thereby concentrating the
cache candidate data sets on a few cachable volumes.
The cache volume creation module tries to achieve the
maximum performance improvement with the minimum
number of moves.
Z00~36
18
Cache Volume Cre~tion Overview
The way that the cache volume creation module 302
operates is to first identify which data sets on the
subsystem need to be cached. Those data sets that are
good cache candidates are classified as qood, and
those which are bad cache candidates are classified
as bad. Based on the classification of the data sets,
the cache volume creation module 302 then determines
which volumes are good cache volumes, which are bad,
and which have good data sets in conflict with bad
data sets. The volumes are classified as qood, bad,
and conflict respectively. The number of cache
volumes required is then determined as are the data
set moves needed to create this number of cache
volumes. Then the conflict volumes are processed and
the good data sets are separated from the bad data
sets. Finally, the cache volume creation module 302
executes a last pass which checks if it can make
significant improvements to cache candidate volumes
with one or two moves.
Cache Volume Creation Module
The cache volume creation module 302 contains the
function entry point that is called by the scheduler
lO9. The cache volume creation module 302 then
executes the sub-modules 303-308 as shown in Figure
3. When the cache volume creation module 302 exits
it calls operating system 102 to execute the list of
recommended data set movements and update the
knowledge database 108 with the recommendations.
On entry, a pointer to a subsystem record for the
subsystem being analyzed is passed. This record
contains the subsystem information and also contains
a pointer to the list of volume records, one for each
;~0~36~
19
volume of the subsystem. Each of the volume records
contains the volume information and a pointer to the
list of data set records, one for each data set on the
volume. Each data set record contains information on
data set activity.
Initialize
The initialize module 303 performs the
initialization required by the cache volume creation
function including any data format conversion that
needs to take place. The initialize module 303 also
looks for data sets with moves scheduled for them.
At the time the data is collected for analysis, these
data set moves had not been completed, but by the time
any moves recommended by the current analysis are
performed, these moves will have been carried out.
This means that the state of the machine is different
from the time the data was collected to the time the
data set move recommendations are implemented and so
data set move recommendations are based on the system
as it would look if the moves already scheduled had
been performed. The initialize module 303 alters the
input data to reflect this state of the system.
Determine Cache Candidates
The purpose of the determine cache candidates
module 304 is to determine which data sets on the
subsystem to place on cached volumes. First the
routine Classify Data Sets 311 is called which
classifies the data sets as good, bad, or inactive
based on how well they will perform if they are
cached. Then the routine Determine Cutoff 312 is
called. This routine determines if there is too much
cache candidate data on the subsystem for the size of
2006~336
the cache and if so determines a cutoff that reduces
the amount of cache candidate data down to the best
data sets the size of cache will cope with. The
Reclas~ify Data Sets 313 routine then reclassifies the
cache candidate data sets that fall below the cutoff
and flags them for possible movement to another cache
subsystem.
Classify Data 8ets
The classify data set routine 311 classifies the
data sets on the subsystem as qood cache candidates,
bad cache candidates, or inactive data sets. The
classification is done by some simple rules of thumb
based on the activity, the read percent, and the LORM
of the data set. (The LORM is the Locality Of
Reference Measure and is equal to the number of I/Os
per second divided by the number of unique tracks
accessed over the monitored period.)
As well as classifying the data sets as being
good or bad cache candidates, the classify data sets
module also classifies the good data sets into classes
which reflect the degree to which the data set is
suitable for caching. Thus, in selecting data sets
to cache for the given cache size, the classify data
sets module attempts to include the better cache
candidates.
Rules
The rules used for classifying the data sets are
shown below, where x, y and z are variables that are
selected as optional for the particular device
characteristics and the user's needs.
(1)
if 1. there was no activity on the data set
then
2006~3~6
21
the data set is inactive
(2)
if 1. the read% of the data set > x and
2. the LORM > y and
3. the I/O rate > z per second
then
the data set is good
(3)
if 1. (a) the read% < x or
(b) the LORM < y or
(c) the I/O rate < z per second and
2. there was some activity on the data set
then
the set is bad
Determine Cutoff
The determine cutoff routine 312 is responsible
for deciding which cache candidates identified by the
classify data sets routine 311 should be cached based
on the size of the cache. Starting with the best
cache candidates, and based on attributes of the data
sets such as activity, read percent, and LORM, the
determine cutoff routine 312 works its way down the
list of cache candidates until a sufficient number of
cache candidates are selected for the available cache
size. This is where the determine cutoff routine 312
determines there should be a cutoff with those data
sets above the cutoff being cached. In deciding where
the cutoff is, the determine cutoff routine 312 also
looks at the currently cached data and the read hit
rate of the cache and devices. This way, if the cache
is being overloaded, some data sets can be moved off
the cache volumes and flagged for possible movement
off the subsystem. Similarly, if the cache is
performing better than expected and can handle more
data, some more cache candidates are moved onto the
cache volumes. If the determine cutoff routine 312
Z~:)6~ 6
22
determines that the given cache size can handle more
than the amount of cachable data that is available on
the subsystem, the determine cutoff routine 312 flags
the subsystem as having an under-utilized cache.
Reclassify Dat~ 8et~
Using the cutoff determined by the determine
cutoff routine 312, the reclassify data sets routine
313 reclassifies the data sets which fall below the
cutoff as bad data sets and flags them for possible
lo movement to another cache subsystem.
ClassifY Volumes
The classify volumes routine 305 classifies the
volumes of the subsystem as being qood, bad, or
conflict volumes based on whether the volume is a good
cache candidate, a bad cache candidate, or whether
there is good cache data in conflict with bad cache
data on the volume. The classify volumes routine 305
uses the classifications of the data sets to determine
the classification of the volume. If the volume
predominantly consists of good data sets, then the
volume is classified as qood. If the volume consists
totally of bad data sets, then the volume is
classified as a bad volume. If there is good data on
the volume but there is also a significant amount of
bad data, then there is a conflict between the good
and bad data and the volume is classified as a
conflict volume.
Rule~
The rules used to perform the volume
classification are shown below where n is a
predetermined threshold selected as a function of the
2~(~fi~336
-
23
characteristics of the device being monitored and the
needs of the user.
(1)
if 1. there are no good data sets on the volume
then
the volume is bad
(2)
if 1. there is good data on the volume and
2. the amount of bad data accounts for
more than n of the total activity
of the volume
then
the volume is a conflict volume
(3)
if 1. there is good data on the volume and
2. the amount of bad data accounts for
less than n of the total volume
activity
then
the volume is good
Create Cache Volumes
The create cache volumes routine 306 is
responsible for deciding how many cache volumes there
should be on the subsystem and creating them if there
are not enough qood volumes already in existence.
First the determine number of cache volumes routine
314 is used to determine the desired number of cache
volumes based on a predetermined device activity level
and space constraints. The number of desired cache
volumes is compared with the number of good cache
volumes in existence to determine how many more good
cache volumes must be created. The Determine
Resolution gtrategy routine 315 decides how to resolve
the conflicts on conflict volumes. The determine
resolution strategy routine 315 finds volume that will
end up as good volumes by looking for volumes where
the resolution strategy is to move the bad data sets
Z0~6t336
-
24
off the volume. This is how the determine resolution
strategy routine 315 creates the required number of
cachable volumes. If there are not enough conflict
volumes with a resolution strategy of moving the bad
data sets, then the Create Cache Volume~ routine 306
finds volumes which require the least number of moves
to turn into good volumes. The Data ~et Placement
routine 316 is used to decide where to relocate data
sets needing movement to create cachable volumes.
Determine Number of Cache Volumes
The determine number of cache volumes routine 314
decides how many cache volumes there should be on the
subsystem. It bases this decision on the space
requirements of the identified cache candidates, the
activity of the cache and non-cache candidates, the
anticipated read hit rate of the cache and the
devices, and the amount of space on the whole
subsystem. The aim is to have enough cache volumes
that device busy queuing is not a problem and space
requirements for cache data are met, as well as trying
to minimize device contention problems and space
problems on non-cache volumes.
Resolve Conflict~
The resolve conflicts routine 307 processes the
remaining conflict volumes by resolving the conflicts
on the volumes. A conflict is where there is cache
data and a significant amount of non-cache data
residing on the same volume. The volume can't be
cached because the bad data would degrade the
performance of the cache too much. The resolve
conflicts routine 307 decides on a resolution
strategy, and then implements the strategy.
~00~ 6
`_
The resolution strategy is determined by the
Determine Re~olution 8trategy routine 317 and can be
one of
1. move good,
2. move bad, or
3. salvage qood.
That is, in order to resolve the conflict, one
of the following actions can be selected
1. move the good data sets to good volumes,
2. move the bad data sets to bad volumes, or
3. salvage as many good data sets as we can by
moving the movable good data sets with
significant activity to good volumes.
Once the resolution strategy has been directed,
the Data Set Placement routine 317 is used to decide
where to move the identified data sets.
Determine Resolution gtrategy Routine
Given a conflict volume, the best way of
eliminating the conflict must be determined, that is,
the best way to separate the good and bad data from
each other. The good data sets can be relocated off
the volume, the bad data sets can be relocated off the
volume, or as many good data sets as possible can be
salvaged.
Salvaging the good data sets is used as a last
resort where it is not possible to move the bad data
sets off the volume due, perhaps, to immovable data
sets or the number of moves involved and it is
similarly not possible to move all of the good data
sets off the volume. Under these circumstances, good
data sets are salvaged by moving those good data sets
that have significant activity and which are movable.
In referencing bad data sets on a conflict
volume, what is actually meant is a selected subset
of the bad data sets. A good volume can still have
Z006836
26
bad data sets on it as long as the activity due to the
bad data sets is less than n of the volume activity.
Thus, in order to turn a conflict volume into a good
volume enough bad data sets must be moved to reduce
the bad activity to this level. The selected bad data
sets are then those which if moved off the volume
would reduce the amount of bad activity to below 20%
of the volume activity.
The factors which influence the decision about
which resolution strategy to use are listed below.
1. immovable bad
2. inmo~rable qood
3. w~ ng to move bad
4. w- l-ng to move qood
5. wh ch needs more moves
6. wh ch has more activity
7. qood data sufficiently active
8. which ls more immovable
9. some bad is totally immovable
10. some qood is totally immovable
11. enough unwilling qood
Each of these factors is represented as an
attribute of the conflict volume whose value is
derived from input data and other attributes. The
derivation of the value of the above attributes and
the attributes from which these values are obtained
are explained in detail in later sections.
Rules
The rules which use these factors to decide on
a strategy to use to resolve the conflict are shown
below.
(1)
If l. there are no immovable bad data sets
and
2. there are some totally immovable bad
data sets
Then the resolution strategy is ERROR
;~0~ 336
27
(2)
If 1. there are no immovable good data sets
and
2. there are some totally immovable good
data sets
Then the resolution strategy is ERROR
(3)
If 1. the good and bad data are equally
immovable and
2. there are some immovable good data sets
and
3. there are no immovable bad data sets
Then the resolution strategy is ERROR
(4)
If 1. the good and bad data are equally
immovable and
2. there are some immovable bad data sets
and
3. there are no immovable good data sets
Then the resolution strategy is ERROR
(5)
If 1. the good and bad data are equally
immovable and
2. there are some totally immovable good
data sets and
3. there are no totally immovable bad data
sets
Then the resolution strategy is ERROR
(6)
If 1. the good and bad data are equally
immovable and
2. there are some totally immovable bad
data sets and
3. there are no totally immovable good
data sets
Then the resolution strategy is ERROR
(7)
If 1. the good data is more immovable and
2. there are no immovable good data sets
Then the resolution strategy is ERROR
(8)
If 1. the bad data is more immovable and
2. there are no immovable bad data sets
Then the resolution strategy is ERROR
ZOOfi~3~6
28
If 1. the good data is more immovable and
2. there are some totally immovable bad
data sets
Then the resolution strategy is ERROR
(10)
If 1. the bad data is more immovable and
2. there are some totally immovable good
data sets
Then the resolution strategy is ERROR
( 11)
If 1. we are willing to move all of the good
data sets and
2. we are not willing to move the bad data
and
3. there are no immovable good data sets
Then the resolution strategy is MOVE_GOOD
(12)
If 1. we are willing to move the bad data
and
2. we are not willing to move all of the
good data sets and
3. there are no immovable bad data sets
and
4. there is enough unwilling good data to
justify any extra moves required to
move the bad data
Then the resolution strategy is MOVE_BAD
(13)
If 1. we are willing to move all of the
good data sets and
2. we are not willing to move the bad data
and
3. there are some immovable good data sets
and
4. the good data is sufficiently active
and
5. there are no totally immovable good
data sets
40Then the resolution strategy is MOVE_GOOD
(14)
If 1. we are willing to move the bad data
and
2. we are not willing to move all of the
45good data sets and
3. there are some immovable bad data sets
and
2()0~36
4. the good data is sufficiently active
and
5. there is enough unwilling good data to
justify any extra moves required to
move the bad data and
6. there are no totally immovable bad data
sets
Then the resolution strategy is MOVE_BAD
(15)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are no immovable good data sets
and
4. there are no immovable bad data sets
and
5. the good data requires more moves
Then the resolution strategy is MOVE_BAD
(16)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are no immovable good data sets
and
4. there are no immovable bad data sets
and
5. the bad data requires more moves
Then the resolution strategy is MOVE_GOOD
(17)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are no immovable good data sets
and
4. there are no immovable bad data sets
and
5. the good and bad data require the
same number of moves and
6. the good data has more activity
Then the resolution strategy is MOVE_BAD
(18)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
Z00~i~36
_~ 30
3. there are no immovable good data sets
and
4. there are no immovable bad data sets
and
5. the good and bad data require the same
number of moves and
6. the bad data has higher activity
Then the resolution strategy is MOVE_GOOD
(19)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are no immovable bad data sets
Then the resolution strategy is MOVE_BAD
(20)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are no immovable good data sets
and
4. there are some immovable bad data sets
Then the resolution strategy is MOVE_GOOD
(21)
If 1. we are willing to move all of the
good data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the good data is more immovable than
the bad data and
7. there are no totally immovable bad data
sets
Then the resolution strategy is MOVE_BAD
(22)
If 1. we are willing to move all of the good
data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
20~6~36
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the bad data is more immovable than
the good data and
7. there are no totally immovable good
data sets
10Then the resolution strategy is MOVE_GOOD
(23)
If 1. we are willing to move all of the good
data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the good and bad data are equally
immovable and
7. there are no totally immovable good
25data sets and
8. there are no totally immovable bad data
sets and
9. the good data requires more move
Then the resolution strategy is MOVE_BAD
(24)
If 1. we are willing to move all of the good
data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the good and bad data are equally
immovable and
7. there are no totally immovable good
data sets and
8. there are no totally immovable bad data
sets and
9. the bad data requires more moves
Then the resolution strategy is MOVE_GOOD
ZO~t~836
(25)
If 1. we are willing to move all of the good
data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the good and bad data are equally
immovable and
7. there are no totally immovable good
15data sets and
8. there are no totally immovable bad data
sets and
9. the good and bad data require the same
number of moves and
10. the good data has higher activity
Then the resolution strategy is MOVE_BAD
(26)
If 1. we are willing to move all of the good
data sets and
2. we are willing to move the bad data
and
3. there are some immovable good data sets
and
4. there are some immovable bad data sets
and
5. the good data is sufficiently active
and
6. the good and bad data are equally
immovable and
7. there are no totally immovable good
data sets and
8. there are no totally immovable bad data
sets and
9. the good and bad data require the same
40number of moves and
10. the bad data has higher activity
Then the resolution strategy is MOVE_GOOD
(27)
If 1. we are not willing to move all of the
45good data sets and
2. we are not willing to move the bad data
Then the resolution strategy is SALVAGE_GOOD
(28)
If 1. there are some totally immovable good
zoo~
data sets and
2. there are some totally immovable bad
data sets
Then the resolution strategy is SALVAGE_GOOD
(29)
If 1. there are some immovable good data sets
and
2. there are some immovable bad data sets
and
3. the good data is not sufficiently
active
Then the resolution strategy is SALVAGE_GOOD
(30)
If 1. we are willing to move all of the good
data sets and
2. we are not willing to move the bad data
and
3. there are some immovable good data sets
and
4. ( there are some totally immovable
good data sets or
5. the good data is not sufficiently
active )
Then the resolution strategy is SALVAGE_GOOD
(31)
If 1. we are willing to move the bad data
and
2. we are not willing to move all of the
good data sets and
3. there are no immovable bad data sets
and
4. there is not enough unwilling good data
to justify any extra moves required to
move the bad data
35Then the resolution strategy is SALVAGE_GOOD
(32)
If 1. we are willing to move the bad data
and
2. we are not willing to move all of the
40good data sets and
3. there are some immovable bad data sets
and
4. ( the good data is not sufficiently
active or
5. there is not enough unwilling good
data to justify any extra moves
required to move the bad data or
6. there are some totally immovable
20~)6~336
34
bad data sets )
Then the resolution strategy is SALVAGE_GOOD
I ovable Bad
This attribute is true if there are any bad data
sets (in the subset already selected) which are
immovable.
Immovable Good
This attribute is true if any of the good data
sets on the volume are immovable.
Willing to Move Bad
This attribute is true if we are willing to move
the bad data sets off the volume. That is, it is true
if the number of moves required to move the bad data
sets is less than or equal to the number of moves we
are willing to make.
The value of this attribute depends on the value
of two other attributes:
moves willing, and
number of moves to move bad.
Willing to Move Good
This attribute is true if the decision is made
to move all of the good data sets off the volume.
With the bad data sets, the number of moves it took
to move the data sets is compared with a selected
number of moves to make to determine whether to move
the bad data sets. With the good data sets though,
the data sets may be move candidates based on the
number of moves selected, but not selected to move
because there are some good data sets which aren't
worth moving due to their low activity. So, this
attribute is true if all of the good data sets are
Z0~6~336
-
worth moving and the number of moves required is less
than or equal to a selected number of moves.
This attribute depends on the value of these
attributes:
moves willing, and
number of moves to move qood.
Which Needs More Moves
The value of this attribute indicates whether it
takes more moves to move the good data sets off the
volume or whether it takes more moves to move the bad
data sets. The values it can take are:
qood,
bad, or
same.
This attribute depends on the value of the
following attributes:
number of moves to move bad, and
number of moves to move qood.
Which Has More Activity
The value of this attribute indicates whether
the activity of the good data is greater than or equal
to the bad data or whether the bad data has more
activity than the good data. The values it can take
are:
good, or
bad.
Good Data Sufficiently Active
This attribute is true if the combined activity
of the good data is sufficiently active to consider
moving immovable data sets. This decision is based
on the activity of the good data on the volume
relative to the total amount of good activity on the
ZO~ 36
_ 36
subsystem and also on the importance of obtaining more
cache data.
The value of this attribute depends on the
following attributes:
activity classification, and
importance of obtaining more cache data.
Rules
The rules used to determine whether the good data
is sufficiently active are shown below.
(1)
if 1. the good data has moderate activity and
2. it is very important to obtain more
cache data
then
the good data is sufficiently active
(2)
if 1. the good data has a high activity and
2. it is very important to obtain more
cache data
then
the good data has sufficient activity
(3)
if 1. the good data has very high activity
then
the good data has sufficient activity
Which is More Immovable
The value of this attribute indicates which of
the good and bad data is more immovable than the
other. At the moment, one type of data is more
immovable than the other if the most immovable data
set of the first type is more immovable than the most
immovable data set of the other type. The values the
attribute can have are:
qood,
bad, or
same.
~0C~ 6
_ 37
Some Bad Totally Immovable
This attribute is true if there is at least one
bad data set which is totally immovable and must not
be moved under any circumstances. For example VTOCs
must never be moved for any reason.
Some Good Totally Immovable
This attribute is true if some of the good data
sets are totally immovable.
Enough Unwilling Good
When there are good data sets that are not move
candidates (because their activity is not high
enough), this parameter indicates whether there are
enough of these unwilling data sets to justify any
extra moves it might take to move the bad data sets
off the volume. The alternative is to try and move
the good data sets that we are willing to move and
leave the unwilling good data sets behind to end up
on a bad volume. This attribute is true if the
combined activity of the unwilling good data sets is
enough to justify any extra moves required in moving
the bad data sets off the conflict volume.
Number of Moves to Move Bad
The value of this attribute indicates the number
of moves it will take to move the bad data sets off
the conflict volume. The number of moves is
determined by the Data 8et Placement module.
Number of Moves to Move Good
This attribute gives the number of moves required
to move all of the good data sets off the conflict
~006~336
-
38
volume. The number of moves is determined by the Data
Set Placement module.
Moves ~illin~
When making movement recommendations, the
procedure is to minimize the number of moves. For
each conflict volume there is a number of moves that
can be made based, principally, on the activity of the
good data relative to the total amount of good data
on the subsystem. So, where there is a lot of good
data, more moves are desireable and where there is
only a small amount of good data fewer moves are
desireable. As well as the activity of the good
data, the value of the attribute importance of
obtaining more cache data is scrutinized. If it is
important to put more cache data on the cache volumes
then the number of moves permitted is correspondingly
higher.
Rules
The rules used for determining the number of
moves are shown below.
(1)
if 1. the activity of the good data set is X%
of the total amount of good data on the
subsystem and
2. it is not important to obtain more
cache data
then
the number of moves we are willing to make
for this data set is X/5
(2)
if 1. the activity of the good data set is X%
of the total amount of good data on the
subsystem and
2. it is important to obtain more cache
data
then
~006~
_ 39
the number of moves we are willing to make
is l.S * X/5
(3)
if 1. the activity of the good data set is X%
of the total amount of good data on the
subsystem and
2. it is very important to obtain more
cache data
then
the number of moves permitted
is 2 * X/5
Importance of Obtai ni ng More Cache Data
This attribute reflects how important it is to
obtain more cache data. Of course, it is always
important to obtain more cache data but this attribute
indicates whether it is more important than usual or
not. The decision about whether it is important to
obtain more cache data is based on the amount of cache
data so far obtained and the amount anticipated. The
possible values for this attribute are:
not important,
important, or
very important.
The way that this module measures the amount of
good data it anticipates will end up on good volumes
is to measure the amount of good data on easilY
resolved conflict volumes and add this to the amount
of good data already residing on good volumes. This
method gives a very rough measure that is used to
determine whether to work harder at getting more good
data onto cache volumes.
The concept of easily resolved conflicts is used
to try and predict how much good data will end up
cachable. The simplistic approach taken is to assume
that the amount of good data that is not already
cachable but that will end up cachable is about the
20~ 6
-
same as the amount of good data currently residing on
easily resolved conflict volumes.
Rules
The rules that are used to determine the
importance of obtaining more cache data are shown
below where r and s are user selected variables.
(1)
if 1. the amount of good data on good volumes
is < r of total good measure by
activity and
2. there is < r total good data (by
activity) on easily resolved conflict
volumes
then
it is very important to get more cache data
(2)
if 1. if between r and s of the good data
(by activity) is cachable and
2. there is < r total good on easily
resolved
then
it is important to get more cache data
(3)
if 1. more than s total good data is
cachable
then
it is not important to get more cache data
(4)
if 1. there is < r total good data on good
volumes and
2. there is between r and s total good
on easily resolved conflict volumes
then
it is important to get more cache data
Ea~ilY Resolved Conflicts
In order to determine the importance of obtaining
more cache data it is necessary to measure the amount
of good data on easily resolved conflict volumes. A
conflict volume is considered to be easily resolved
2()06~ 6
-
41
if:
1. there are no immovable bad data sets, and
2. the number of moves we are willing to make
for the good data is greater than or equal
to the number of bad data sets that would
have to be moved to turn the volume into a
good volume.
Activity Classification
This module is used to classify the activity of
some good data. The classification is one of:
low,
moderate,
high, or
very high.
The classification is based on the percentage of
the total amount of good activity on the subsystem
attributable to the good data being classified.
Rules
The rules used are shown below where t, u and v
are preselected variables that are a function of
device characteristics and user needs.
(1)
if 1. the activity of the good data is >= t
of total good activity
then
the activity is very high
(2)
if 1. the activity of the good data is >= u
of total good and
2. the activity is <= t total good
then
the activity is high
(3)
if 1. the activity is >= v total good and
2. the activity is < u total good
then
the activity is moderate
~006836
-
42
(4)
if 1. the good data has some activity and
2. the activity is ~ v total good
then
the activity is low
8alvage Good Data
During conflict resolution, the preferred options
for resolving the conflict are to move all of the
selected bad data sets off the volume or to move all
of the good data sets off the volume. However, due
to immovable data sets or the activity of the good
data not being high enough to warrant the number of
moves needed, it is possible that neither of these two
options can be recommended. In this case the result
of the resolution process is to recommend that the
good data on the volume be salvaged. The salvage
operation consists of ex~;ning each good data set on
its own to determine whether it is worth moving to a
good volume, thus making it cachable.
If a good data set is not immovable, then it is
worth salvaging if the number of moves permitted is
enough to move it. If a good data set is immovable
but not totally immovable, then not only does the
number of moves permitted have to be enough to move
it, but it has to have sufficient activity to warrant
its movement off the volume. A totally immovable good
data set can never be salvaged.
Rules
The above conditions for salvaging a good data
set can be expressed as the rules shown below.
(1)
if 1. we are willing to move the good data
set and
2. the data set is not immovable
then
~0~ 36
43
move the data set
(2)
if 1. we are willing to move the good data
set and
2. the data set is immovable and
3. the data set has sufficient activity
then
move the data set
Data 8et Placement
The data set placement routine 316 is used to
decide where to place data sets that have been
identified for movement. It attempts to recommend
placement of a data set on the least active volume of
the same classification on which it will fit. If
there are no matching volumes on which the data set
will fit, the data set placement routine 316 tries to
find another data set on a matching volume that, when
moved to another matching volume, frees enough space
for the original data set. As a last resort the data
set placement routine 316 considers moving the data
set to another conflict volume for processing at a
later date when hopefully it is easier to place.
When a place has been found for a data set to be
moved to, a recommendation is generated and placed on
the end of the list of current recommendations by
calling the Generate Recommendations routine 318.
As well as actually placing data sets, the data
set placement routine 316 is used during the conflict
resolution process to determine how many moves are
necessary to move a given set of data sets off a
volume. This information is then used when deciding
which of the good and bad data need more moves to be
moved off a conflict volume.
~OOti~336
44
Generate Recommendations
This routine takes data set movements and adds
them to the list of current recommendations.
Last Pass Optimisation
Just before the cache volume creation routine 302
finishes its analysis, the last pass optimization
routine 308 is called in an attempt to find good
volumes where one or two movements will result in a
dramatic performance increase. The last pass
optimization routine 308 looks at all of the good or
cache candidate volumes and identifies where there are
still one or two moderately high activity bad data
sets that could be moved off the volume to increase
the cache suitability of the volume.
If there is a bad data set on a good volume that
accounts for more than a certain percentage of the
total volume activity and can be moved off with only
one move, then the movement of the data set off the
volume is recommended (309). Similarly, if there is
a bad data set on a good volume that accounts for more
than a predetermined percentage of the total volume
activity and it can be moved, then it is recommended
(309) for movement off the volume.
DASD Performance OPtimizer
The DASD Performance Optimizer (DPO) module 106
is one of the tuning analysis routines of the computer
system memory performance improvement apparatus 103.
The DASD performance optimizer module 106 uses Expert
System and artificial intelligence techniques to
analyze the performance of the DASD devices in an
installation and make recommendations of data set
movements that improve I/O performance.
Z00&t336
_ 45
The purpose of the DASD performance optimizer
module 106 is to make data set movement
recommendations with the aim of improving the
performance and utilization of the DASD devices. The
module tries to minimize the number of movements
recommended. The recommendations are based on data
that has been collected as a result of monitoring the
I/O activity and characteristics of the devices. The
module works in such a way as to prevent problems
occurring rather than waiting for problems to occur
before doing anything about them. The DASD
performance optimizer module 106 identifies devices
and subsystems where problems could occur or where
there is already a problem, and makes a few data set
movement recommendations that will rectify the
situation. In this way, data sets are moved from
over-utilized volumes and subsystems to under-utilized
volumes and subsystems. The DPO module 402 also
identifies where there are cache candidate data sets
that could be moved to subsystems whose cache is being
under-utilized. If necessary, it also moves data sets
off cache subsystems when the cache on the subsystem
is over-loaded.
When any initial setup that is required has been
done by operating system 102, it calls the DPO module
402 as a subroutine passing any information the module
requires in a specified format. When the DPO module
402 terminates, it returns to the operating system 102
which performs any cleaning up that is needed before
exiting.
DASD Performance Optimizer
After some initialization, the DASD performance
optimizer module 402 tries to reorganize data sets
~OO~ 36
46
that have been flagged for movement by the Cache
Volume Creation module 105 so that they end up under
cache controllers. The subsystems and volumes that
have problems or potential problems are identified.
Then sets of data sets on these subsystems and volumes
are identified, whose movement would rectify the
situation. Each of the over-utilizations identified
are then processed ordered by their priority and three
alternative data set combinations are selected for
each. The processing involves choosing the best of
the sets of data sets, deciding where to move them,
and generating appropriate recommendations.
The DASD performance optimizer module 402
controls the high level flow of execution and it calls
operating system 102 to execute the list of data set
move recommendations and to write these
recommendations to the knowledge database 108.
Initialize
The initialize routine 404 performs any
initialization required including reformatting data
if this is needed. This initialization process
includes reading information from the database to
create internal records for the data storage complex,
each subsystem in the complex and each volume in each
subsystem for which a record in the database exists.
The initialization routine 404 also obtains pool and
host sharing information about each of the volumes
stored in the data storage system and creates objects
for each of the compatibility classes that exist. A
compatibility class is a combination user pool, set
of sharing hosts and cachability type. If any moves
are scheduled for data sets, then initialize routine
404 alters the data to reflect the state the system
;~006S~3~
`_
47
would be in if the movements had been executed. The
reason for doing this is that any recommended
movements are executed after any moves that have been
already scheduled and so data set move analysis is
based on the assumption that the moves have been
carried out.
Inter-Subqystem Cache Reor~anization
The Inter-Subsystem Cache Reorganization routine
405 exists to support the movement of cache candidate
data sets across subsystems. The cache volume
creation function identifies where there are under-
utilized caches and also where there are cache
candidate data sets that can't be cached due to a lack
of cache space on the subsystem. DASD performance
optimizer module 402 then looks for under-utilized
caches and tries to find cache candidate data sets
that have been flagged for movement that it can move
to these cache subsystems.
The DASD performance optimizer module 402 uses
the cache suitability classification that cache volume
creation 302 gives each cache candidate data set to
decide which data sets should be moved to which
subsystems. If the DASD performance optimizer 402
finds that there is a lot of good cache candidate data
that is uncached and that there is not enough cache
capacity on any of the cached subsystems, then this
module writes out a message to user recommending that
the user buy some more cache.
Identification
The job of the identification module 406 is to
find subsystems and volumes which are over-utilized.
A volume or subsystem becomes a problem when the
20~:)fi~336
48
amount of activity is such that the amount of queuing
for the device or for a path is unacceptable. Volumes
or subsystems are identified as over-utilized when
they reach an activity level above some safety margin
below the unacceptable level. This safety margin
allows actual problems to be corrected as well as
preventing problems from arising by detecting places
that potentially could become problems.
The detection of these over-utilizations is based
on the percent busy of a device or path, and the
number of paths. The percent busy measures are
derived from the number of I/Os per second, the
average data size transferred, the channel speed, the
device service time, and the number of paths. As well
as identifying which devices and subsystems are over-
utilized this module also decides how much data needs
to be moved off the device or subsystem. This is
represented by an amount of path time per second or
device time per second that we need to reduce by.
RuleQ
Some examples of the sort of rules that might be
used for determining whether a device or subsystem is
over-utilized are shown below where a and b are
predetermined variables.
(1)
if a subsystem is not a cache subsystem and
it is a subsystem of 8380 class devices and
there are two paths and
there is not actuator level buffering (ALB)
30 and
the paths are more than a busy
then
the subsystem is over-utilized
(2)
if a device is on a non-cache subsystem and
it is an 8380 class device and
the device is more than b busy
X006~;~6
49
then
the device is over-utilized
(3)
if a subsystem is a cache subsystem and
there are two paths and
the devices are 8380 class devices and
there is not ALB and
the paths are more than a busy (due to
staging and I/O path time)
then
the paths are over-utilized
Identification module 406, in addition to
identifying over-utilized subsystems and volumes,
calculates the degree of over-utilization of each of
these over-utilized subsystems. For example, where
a particular subsystem has been identified as over-
utilized, identification module 406 determines the
amount of activity handled by this subsystem by for
example adding up all of the input and output
activities to the various volumes that constitute this
subsystem. The identification module 406 then
determines the capacity of the subsystem for extra
activity. That is, it determines the number of extra
input and output operations that the subsystem could
handle without becoming overloaded. Once this
calculation is completed, identification module 406
assigns a classification attribute to this particular
subsystem which attribute is indicative of the degree
ov over-utilization of the subsystem. This procedure
is followed for each subsystem in the entire system
that has been identified as over-utilized and
similarly for each volume that has been identified as
over-utilized.
8electing Alternativeq
Once the over-utilizations have been identified,
the selecting alternatives routine 407 finds
- Z00~8;~6
combinations of data sets that when moved off the
volume or subsystem will solve the problem. The
selecting alternatives routine 407 selects three
alternate combinations, known as alternatives for each
over-utilized volume or subsystem. Each alternative
is a set of data sets whose movement off the volume
or subsystem will rectify the over-utilization. The
reason for selecting three alternatives is so that,
the over-utilization process, if there is a shortage
of locations to place data sets, then there are three
different alternatives, one of which will be able to
be moved. The selection of the alternatives is done
by first generating candidate alternatives via the
Candidate Generation routine 411, and then selecting
the best three candidates via the Candidate Selection
routine 412.
Candidate Generation
The candidate generation routine 411 generates
candidate combinations of data sets whose combined
activity is enough to rectify the over-utilization if
they were moved off the volume or subsystem.
Candidate generation routine 411 determines the
availability of space and activity capacity for each
of the volumes in the compatibility class to which an
identified over-utilized volume is assigned. The
availability measure is a variable that indicates how
much available space and activity capacity is
available on each volume. Greater significance is
given to the occurrence of both space and activity
capacity on a volume rather than the availability of
one or the other. In addition, the same calculation
is performed to identify the possibility of moving
data sets from one subsystem to another subsystem.
200~ 6
-
51
This is the case where all of the volumes in the
compatibility class reside on more than one subsystem
in the system. Therefore, relocating a data set to
a different subsystem maintains the data set within
the same compatibility class but reduces the level of
utilization of the subsystem that initially had the
data set stored thereon. In order to reduce the
number of data sets that have to be considered, a
threshold is calculated and data sets with an activity
below the threshold are ignored. When generating the
combinations, immovable data sets are excluded and the
combined activity of the data sets are selected close
to that required for movement. In selecting
candidates for fixing a subsystem problem, candidates
are identified that would also solve any volume
problems there may be within the subsystem.
Candidate 8election
The candidate selection routine 412 selects three
candidates to be the alternatives for a volume or
subsystem. In performing this operation, candidate
selection routine 412 calculates a measure of the
difficulty of placement of a data set on another
volume. This measure indicates the anticipated
difficulty in moving the data set from its present
location on a particular over-utilized volume to
another candidate volume that is not over-utilized and
yet in the same compatibility class. In investigating
the placement of a data set, the candidate selection
routine 412 determines whether the destination volume
is compatible with the source volume. That is, the
destination volume must be shared by the same set of
hosts as the source volume, and must also be in the
same user defined pools. In addition, candidate
2006~336
52
selection routine 412 attempts to move non-cache data
sets to non-cache volumes. Apart from finding a
volume compatible with the source volume of the data
set, candidate selection routine 412 must locate a
volume that enough room for the data set and one which
has enough capacity for extra activity resultant from
the relocation of the data set. One of the goals used
in the selection process is to try to minimize the
number of data sets in an alternative because we are
trying to minimize the number of moves that need to
be performed. It is still possible that some of the
candidate alternatives have immovable data sets in
them so another goal of the selection process is to
try not to select alternatives with immovable data
sets. Large data sets are excluded from alternatives
and to select alternatives for a subsystem that solve
any volume problems within the subsystem. Another
goal is trying to make the three alternatives
different from each other so that being unable to move
one alternative won't necessarily mean another can't
be moved.
Process Over-Utilization
The Process Over-Utilization routine 408
processes all of the over-utilizations by finding the
one with the highest priority and calling the Rectify
Over-Utilization routine 414 to fix it. The intention
is that over-utilizations that are severely impacting
the I/O sub-system or over-utilizations that are
difficult to fix should get a high priority and be
processed first.
200~336
53
Find Highest Over-Utilization
The priority of an over-utilization is based on
the degree of over-utilization and the degree of
difficulty anticipated in rectifying the over-
utilization. The degree of difficulty is based on howmuch space and activity capacity is available and how
much, based on the selected alternatives, a given
over-utilization will need. The degree of over-
utilization is based on the degree to which I/O
performance is degraded for that device or subsystem.
To find the highest priority over-utilization, the
find highest priority over-utilization routine 413
balances the degree of over-utilization and the degree
of difficulty.
Rectify Over-Utilization
Given an over-utilization, the rectify over-
utilization routine 414 fixes the problem by deciding
which set of data sets to move off the volume or
subsystem and where to move them to. This is achieved
by calling the Choo~e Alternative routine 415 to
decide which alternative to move off the volume or
subsystem. The Data Set Placement routine 416 is then
used to decide where to put the data sets of the
alternative.
Choo~e Alternative
The choose alternative routine 415 chooses the
alternative to move by picking the best one for whose
data sets there is space and activity capacity
available.
Z00683~i
54
Data Bet Placement
The data set placement routine 416 decides the
best place to put data sets. Data sets that reside
on shared volumes must only ever be moved to volumes
that are shared by the same hosts and the movement of
a data set should not cause the destination volume to
become over-utilized. If the data set placement
routine 416 finds that there is not enough space
available for the data set it will consider moving one
other data set to make enough room.
Data set placement routine 416 moves the data set
from an over-utilized volume or subsystem to an under-
utilized volume on an under-utilized subsystem. In
the process of deciding where to move the data set to,
one other data set may be moved to make room or to
make enough activity capacity available. The list of
moves done to achieve moving the original data set
stored in memory. There are two different ways that
a data set can be placed. The first is where there
are under-utilized volumes on under-utilized
subsystems which are in the same compatibility class
as the data set and on which the data set will fit.
In this case, data set placement routine 416 places
the data set on the one of the volumes in this
subsystem that has the least activity. The second
manner in which the data set can be placed is the
situation where there are no volumes that satisfy the
above criteria. In this case, data set placement
routine 416 identifies a pair of volumes of the same
type such that if a data set is moved from one of them
to the other, this move will make enough room or
activity capacity available for the original data set.
_ Z00~8~6
8ummarY
The memory performance improvement apparatus
identifies memory performance conflicts, such as a
performance degradation of the computer system data
storage devices due to a plurality of computers in the
computer system attempting to access a common data
storage device. The memory performance improvement
apparatus identifies the performance conflict as well
as the data sets stored on these data storage devices
related to this conflict. Once the data sets involved
in the performance conflict are identified, the memory
performance improvement apparatus determines
alternative memory storage locations for these data
sets and activates various software routines to
transport these conflict data sets to the alternative
data storage locations. The relocation of these
conflict data sets resolves the memory performance
conflict and improves the retrievability of the data
stored on these data storage devices. By performing
the conflict identification and resolution on a
dynamic real time basis, the data storage devices of
the computer system are operated in a more efficient
manner and the retrievability of the data stored on
these data storage devices is significantly improved
without the need for the data management personnel.
The computer system memory performance improvement
apparatus therefore continuously monitors and modifies
the performance of the data storage devices associated
with the computer system.
While a specific embodiment of the invention has
been illustrated, it is expected that those skilled
in the art can and will devise variations of this
system that fall within the scope of the appended
claims.
