Document revision date: 30 March 2001

Two system parameters determine the maximum sizes for the two data structures in the process header as follows:

• GBLPAGES---Defines the size of the global page table. The system working set size as defined by SYSMWCNT must be increased whenever you increase GBLPAGES.
• GBLSECTIONS---Defines the size of the global section table.

Once an image has been created, it can be installed as a permanently shared image. (See the OpenVMS Linker Utility Manual and the OpenVMS System Manager's Manual, Volume 2: Tuning, Monitoring, and Complex Systems). This will save memory whenever there is more than one process actually mapped to the image at a time.

Also, use AUTHORIZE to increase the user's working set characteristics (WSDEF, WSQUO, WSEXTENT) wherever appropriate, to correspond to the expected use of shared code. (Note, however, that this increase does not mean that the actual memory usage will increase. Sharing of code by many users actually decreases the memory requirement.)

3.8.5 Verifying Memory Sharing

If physical memory is especially limited, investigate whether there is much concurrent image activation that results in savings. If you find there is not, there is no reason to employ code sharing. You can use the following procedure to determine if there is active sharing on image sections that have been installed as shareable:

1. Invoke the OpenVMS Install utility (INSTALL) and enter the LIST/FULL command. For example:

   $ INSTALL INSTALL> LIST/FULL LOGINOUT

   
   INSTALL displays information in the following format:

   DISK$AXPVMSRL4:<SYS0.SYSEXE>.EXE LOGINOUT;3 Open Hdr Shar Priv Entry access count = 44 Current / Maximum shared = 3 / 5 Global section count = 2 Privileges = CMKRNL SYSNAM TMPMBX EXQUOTA SYSPRV

2. Observe the values shown for the Current/Maximum shared access counts:
   • The Current value is the current count of concurrent accesses of the known image.
   • The Maximum value is the highest count of concurrent accesses of the image since it became known (installed). This number appears only if the image is installed with the /SHARED qualifier.

The Maximum value should be at least 3 or 4. A lower value indicates that overhead for sharing is excessive.

The scheduler uses a modified round-robin form of scheduling: processes receive a chance to execute on rotating basis, according to process state and priority.

3.9.1 Time Slicing

Each computable process receives a time slice for execution. The time slice equals the system parameter QUANTUM, and rotating the time slices among processes is called time slicing. Once its quantum starts, each process executes until one of the following events occurs:

• A process of higher priority becomes computable
• The process is no longer computable because of a resource wait
• The process itself voluntarily enters a wait state
• The quantum ends

If there is no other computable (COM) process at the same priority ready to execute when the quantum ends, the current process receives another time slice.

3.9.2 Process State

A change in process state causes the scheduler to reexamine which process should be allowed to run.

3.9.3 Process Priority

When required to select the next process for scheduling, the scheduler examines the priorities assigned to all the processes that are computable and selects the process with the highest priority.

Priorities are numbers from 0 to 31.

Processes assigned a priority of 16 or above receive maximum access to the CPU resource (even over system processes) whenever they are computable. These priorities, therefore, are used for real-time processes.

Processes

A process receives a default base priority from:

• /PRIORITY qualifier in the UAF record
• DEFAULT record in the UAF record

A process can change its priority using the following:

• $SETPRI system service.
• DCL command SET PROCESS/PRIORITY to reduce the priority of your process. You need ALTPRI privilege to increase the priority of your process.

A user requires GROUP or WORLD privilege to change the priority of other processes.

Subprocesses and Detached Processes

A subprocess or detached process receives its base priority from:

• $CREPRC system service
• DCL command RUN

If you do not specify a priority, the system uses the priority of the creator.

Batch Jobs

When a batch queue is created, the DCL command INITIALIZE/QUEUE/PRIORITY establishes the default priority for a job.

However, when you submit a job with the DCL command SUBMIT or change characteristics of that job with the DCL command SET QUEUE/ENTRY, you can adjust the priority with the /PRIORITY qualifier.

With either command, increases are permitted only for submitters with the OPER privilege.

3.9.4 Priority Boosting

When real-time processes (those with priorities from 16 to 31) execute, the following conditions apply:

• They never receive a priority boost.
• They do not experience automatic working set adjustments.
• They do not experience quantum-based time slicing.

The system permits real-time processes to run until either they voluntarily enter a wait state or a higher priority real-time process becomes computable.

3.9.6 Tuning

From a tuning standpoint, you have very few controls you can use to influence process scheduling. However, you can modify:

• Base priorities of processes
• Length of time for a quantum

All other aspects of process scheduling are fixed by both the behavior of the scheduler and the characteristics of the work load.

3.9.7 Class Scheduler

The OpenVMS class scheduler allows you to tailor scheduling for particular applications. The class scheduler replaces the OpenVMS scheduler for specific processes. The program SYS$EXAMPLES:CLASS.C allows applications to do class scheduling.

3.9.8 Processor Affinity

You can associate a process or the initial thread of a multithreaded process with a particular processor in an SMP system. Application control is through the system services $PROCESS_AFFINITY and $SET_IMPLICIT_AFFINITY. The command SET PROCESS/AFFINITY allows bits in the affinity mask to be set or cleared individually, in groups, or all at once.

Processor affinity allows you to dedicate a processor to specific activities. You can use it to improve load balancing. You can use it to maximize the chance that a kernel thread will be scheduled on a CPU where its address translations and memory references are more likely to be in cache.

Maximizing the context by binding a running thread to a specific processor often shows throughput improvement that can outweigh the benefits of the symmetric scheduling model. Particularly in larger CPU configurations and higher-performance server applications, the ability to control the distribution of kernel threads throughout the active CPU set has become increasingly important.

This chapter describes tools that help you evaluate the performance of the three major hardware resources---CPU, memory, and disk I/O.

Discussions focus on how the major software components use each hardware resource. The chapter also outlines the measurement, analysis, and possible reallocation of the hardware resources.

You can become knowledgeable about your system's operation if you use MONITOR, ACCOUNTING, and AUTOGEN feedback on a regular basis to capture and analyze certain key data items.

4.1 Prerequisites

It is assumed that your system is a general timesharing system. It is further assumed that you have followed the workload management techniques and installation guidelines described in Chapter 1 and Section 2.4, respectively.

The procedures outlined in this chapter differ from those in Chapters 5 and 10 in the following ways:

• They are designed to help you conduct an evaluation of your system and its resources, rather than to execute an investigation of a specific problem. If you discover problems during an evaluation, refer to the procedures described in Chapters 5 and 10 for further analysis.
• For simplicity, they are less exhaustive, relying on certain rules of thumb to evaluate the major hardware resources and to point out possible deficiencies, but stopping short of pinpointing exact causes.
• They are centered on the use of MONITOR, particularly the summary reports, both standard and multifile.

You should exercise care in selecting the items you want to measure and the frequency with which you capture the data.

If you are overzealous, the consumption of system resources required to collect, store, and analyze the data can distort your picture of the system's work load and capacity.

As you conduct your evaluations, keep the following rules in mind:

• Complete the entire evaluation. It is important to examine all the resources in order to evaluate the system as a whole. A partial examination can lead you to attempt an improvement in an area where it may have minimal effect because more serious problems exist elsewhere.
• Become as familiar as possible with the applications running on your system. Get to know their resource requirements. You can obtain a lot of relevant information from the ACCOUNTING image report shown in Example 4-1. Compaq and third-party software user's guides can also be helpful in identifying resource requirements.
• If you believe that a change in software parameters or hardware configuration can improve performance, execute such a change cautiously, being sure to make only one change at a time. Evaluate the effectiveness of the change before deciding to make it permanent.

  Note When specific values or ranges of values for MONITOR data items are recommended, they are intended only as guidelines and will not be appropriate in all cases.

Image-level accounting is a feature of ACCOUNTING that provides statistics and information on a per-image basis.

By knowing which images are heavy consumers of resources at your site, you can better direct your efforts of controlling them and the resources they consume.

Frequently used images are good candidates for code sharing; whereas images that consume large quantities of various resources can be forced to run in a batch queue where the number of simultaneous processes can be controlled.

4.3.1 Guidelines

You should be judicious in using image-level accounting on your system. Consider the following guidelines when using ACCOUNTING:

• Enable image-level accounting only when you plan to invoke ACCOUNTING to process the information provided in the file SYS$MANAGER:ACCOUNTING.DAT.
• To obtain accounting information only on specific images, install those images using the /ACCOUNTING qualifier with the INSTALL commands ADD or MODIFY.
• Disable image-level accounting once you have collected enough data for your purposes.
• While image activation data can be helpful in performance analysis, it wastes processing time and disk storage if it is collected and never used.

You enable image-level record collection by issuing the DCL command SET ACCOUNTING/ENABLE=IMAGE.

Disable image-level accounting by issuing the DCL command SET ACCOUNTING/DISABLE=IMAGE.

A series of commands like the following generates output similar to that shown in Example 4-1.

$ ACCOUNTING /TYPE=IMAGE /OUTPUT=BYNAM.LIS - _$ /SUMMARY=IMAGE - _$ /REPORT=(PROCESSOR,ELAPSED,DIRECT_IO,FAULTS,RECORDS) $ SORT BYNAM.LIS BYNAM.ORD /KEY=(POS=16,SIZ=13,DESCEND) . . . (Edit BYNAM.ORD to relocate heading lines) . . . $ TYPE BYNAM.ORD

4.3.4 Collecting the Data

Example 4-1 assumes that image-level accounting records have been collected previously.

Example 4-1 Image-Level Accounting Report

From: 8-MAY-1994 11:09(1) To: 8-MAY-1994 17:31 Image name(2) Processor(3) Elapsed Direct(4) Page(5) Total(6) Time Time I/O Faults Records ------------------------------------------------------------------------ EDT 0 00:34:21.34 0 15:51:34.78 5030 132583 390 DTR32 0 00:19:30.94 0 03:17:37.48 7981 83916 12 PASCAL 0 00:15:19.42 0 01:04:19.57 38473 143107 75 MAIL 0 00:10:40.88 1 02:54:02.89 26139 106854 380 LINK 0 00:05:44.41 0 00:23:54.54 7443 57092 111 RTPAD 0 00:04:58.40 0 20:49:19.24 668 8004 72 LOGINOUT 0 00:04:53.98 0 02:01:31.81 2809 67579 893 EMACS 0 00:04:30.40 0 05:25:01.37 420 8461 1 MACRO32 0 00:04:26.22 0 00:14:55.00 1014 34016 46 BLISS32 0 00:03:45.80 0 00:12:58.87 98 32797 8 DIRECTORY 0 00:03:26.20 0 01:22:34.47 1020 27329 275 FORTRAN 0 00:03:13.87 0 00:14:15.08 1157 28003 47 NOTES 0 00:01:39.90 0 02:06:01.95 8011 6272 32 DELETE 0 00:01:37.31 0 00:57:43.31 834 25516 332 TYPE 0 00:01:06.35 0 00:28:58.26 406 14457 173 COPY 0 00:00:57.08 0 00:11:11.40 2197 4943 42 SHOW 0 00:00:56.39 0 00:24:53.22 23 11505 166 ACC 0 00:00:54.43 0 00:03:41.46 132 2007 7 MONITOR 0 00:00:53.91 0 02:37:13.84 159 5649 40 CALENDAR 0 00:00:43.55 0 00:30:15.52 1023 3557 25 PHONE 0 00:00:40.56 0 00:54:59.39 24 1510 33 ERASE 0 00:00:37.88 0 00:03:51.04 105 9873 113 LIBRARIAN 0 00:00:35.58 0 00:03:37.98 1134 10297 62 FAL 0 00:00:34.27 0 00:20:56.63 110 4596 122 SDA 0 00:00:27.34 0 00:09:28.68 52 4797 3 SET 0 00:00:27.02 0 00:02:30.28 160 9447 206 NETSERVER 0 00:00:26.89 0 02:38:17.90 263 10164 407 CDU 0 00:00:24.32 0 00:01:57.67 13 21906 17 VMSHELP 0 00:00:12.83 0 00:05:40.96 121 1943 14 RENAME 0 00:00:09.56 0 00:00:57.44 6 3866 47 SDL 0 00:00:09.55 0 00:01:19.78 11 3158 4 SUBMIT 0 00:00:08.14 0 00:01:08.50 9 2991 28 NCP 0 00:00:07.30 0 00:02:26.20 7 1765 16 QUEMAN 0 00:00:06.44 0 00:01:38.75 201 1561 20

This example shows a report of system resource utilization for the indicated period, summarized by a unique image name, in descending order of CPU utilization. Only the top 34 CPU consumers are shown. (The records could easily have been sorted differently.)

1. Timestamps
   The From: and To: timestamps show that the accounting period ran for about 6 1/2 hours.
   You can specify the /SINCE and /BEFORE qualifiers to select any time period of interest.
2. Image Name
   Most image names are programming languages and operating system utilities, indicating that the report was probably generated in a program-development environment.
3. Processor Time
   Data in this column shows that no single image is by far the highest consumer of the CPU resource. It is therefore unlikely that the installation would benefit significantly by attempting to reduce CPU utilization by any one image.
4. Direct I/O
   In the figures for direct I/O, the two top images are PASCAL and MAIL. One way to compare them is by calculating I/O operations per second. The total elapsed time spent running PASCAL is roughly 3860 seconds, while the time spent running MAIL is a little under 96843 seconds (several people used MAIL all afternoon). Calculated on a time basis, MAIL caused roughly 1/4 to 1/3 of an I/O operation per second, whereas PASCAL caused about 10 operations per second.
   Note that by the same calculation, LINK caused about five I/O operations per second. It would appear that a sequence of PASCAL/LINK commands contributes somewhat to the overall I/O load. One possible approach would be to look at the RMS buffer parameters set by the main PASCAL users. You can find out who used PASCAL and LINK by entering a DCL command:

   $ ACCOUNTING/TYPE=IMAGE/IMAGE=(PASCAL,LINK) - _$ /SUMMARY=(IMAGE,USER)/REPORT=(ELAPSED,DIRECT)

   
   This command selects image accounting records for the PASCAL and LINK images by image name and user name, and requests Elapsed Time and Direct I/O data. You can examine this data to determine whether the users are employing RMS buffers of appropriate sizes. Two fairly large buffers for sequential I/O, each approximately 64 blocks in size, are recommended.

5. Page Faults
   As with direct I/O, page faults are best analyzed on a time basis. One technique is to compute faults-per-10-seconds of processor time and compare the result with the value of the SYSGEN parameter PFRATH. A little arithmetic shows that on a time basis, PASCAL is incurring more than 1555 faults per 10 seconds. Suppose that the value of PFRATH on this system is 120 (120 page faults per 10 seconds of processor time), which is considered typical in most environments. What can you conclude by comparing the two values?
   Whenever a process's page fault rate exceeds the PFRATH value, memory management attempts to increase the process working set, subject to system management quotas, until the fault rate falls below PFRATH. So, if an image's fault rate is persistently greater than PFRATH, it is not obtaining all the memory it needs.
   Clearly, the PASCAL image is causing many more faults per CPU second than would be considered normal for this system. You should, therefore, make an effort to examine the working set limits and working set adjustment policies for the PASCAL users. To lower the PASCAL fault rate, the process working sets must be increased---either by adjusting the appropriate UAF quotas directly or by setting up a PASCAL batch queue with generous working set values.

6. Total Records
   These figures represent the count of activations for images run during the accounting period; in other words, they show each image's relative popularity. You can use this information to ensure that the most popular images are installed (see Section 2.4). For customer applications, you might consider using linking options such as /NOSYSSHR and reassigning PSECT attributes to speed up activations (see the OpenVMS Linker Utility Manual).
   Note that the number of LOGINOUT activations far exceeds that of all other images. This situation could result from a variety of causes, including attempts to breach security, an open terminal line, a runaway batch job, or a large number of network operations. More ACCOUNTING commands would be necessary to determine the exact cause. At this site, it turned out that most of the activations were caused by an open terminal line. The problem was detected by an astute system manager who checked the count of LOGFAIL entries in the accounting file.
   You can also use information in this field to examine the characteristics of the average image activation. That knowledge would be useful if you wanted to determine whether it would be worthwhile to set up a special batch queue.
   For example, the average PASCAL image uses 51 seconds of elapsed time and the average LINK uses 13 seconds. You can therefore infer that the average PASCAL and LINK sequence takes about a minute. This information could help you persuade users of those images to run PASCAL and LINK in batch mode. If, on the other hand, the average time was only 5 seconds, batch processing would probably not be worthwhile.

	privacy and legal statement
6491PRO_004.HTML