21.3.2 How to interpret statistical information about deferred write processing

Information to be referenced
Use the statistics analysis utility (pdstedit) to acquire DAT-format files that contain statistical information about deferred write processing, and then check the following information:
  • Execution time (DWTOTAL, DWTOTALM)
  • Total WRITE time (DWSUM, DWSUMM)
  • Parallel WRITE time (DWPARA, DWPARAM)
  • WRITE unit time
    Minimum (DWMIN, DWMINM)
    Maximum (DWMAX, DWMAXM)
    Average (DWAVG, DWAVGM)
  • WRITE count (DWEXEC)
Organization of this subsection
(1) Execution time (DWTOTAL, DWTOTALM)
(2) Total WRITE time (DWSUM DWSUMM)
(3) Parallel WRITE time (DWPARA DWPARAM)
(4) WRITE unit time minimum (DWMIN DWMINM)
(5) WRITE unit time maximum (DWMAX DWMAXM)
(6) WRITE unit time average (DWAVG DWAVGM)
(7) WRITE count (DWEXEC)

(1) Execution time (DWTOTAL, DWTOTALM)

This is the total time required for deferred write processing. DWTOTAL displays the time in seconds, and DWTOTALM displays only the microseconds portion of the time.

Purpose
This information is collected to determine whether the deferred write processing performance is appropriate.
Evaluating the analysis results
If the deferred write delay message (KFPS02179-I factor code = A01-01) is output, take the following action:
Action to be taken
For details about the action to be taken, see 21.3.1 Tuning procedure.

(2) Total WRITE time (DWSUM DWSUMM)

This is the total time required for writing during deferred write processing. DWSUM displays the time in seconds, and DWSUMM displays only the microseconds portion of the time.

Purpose
This information is collected in order to determine whether tuning the process count when using the facility for parallel writes in deferred write processing had any effect compared to the parallel WRITE time (DWPARA DWPARAM).
Evaluating the analysis results
Check the values of WRITE unit time (maximum) (DWMAX DWMAXM), WRITE unit time (minimum) (DWMIN DWMINM), and WRITE count (DWEXEC).
Action to be taken
Check the values of WRITE unit time (maximum) (DWMAX DWMAXM), WRITE unit time (minimum) (DWMIN DWMINM), and WRITE count (DWEXEC), and take appropriate action.

(3) Parallel WRITE time (DWPARA DWPARAM)

This is the length of time from when a request to process multiple parallel WRITE processes was issued to when completion notifications for all parallel WRITE processes have been received. DWPARA displays the time in seconds, and DWPARAM displays only the microseconds portion of the time. Figure 21-1 shows the concept of parallel WRITE time.

Figure 21-1 Concept of parallel WRITE time

[Figure]

If the facility for parallel writes in deferred write processing is disabled, the parallel WRITE time is 0.

Purpose
This information is collected in order to check the effects of using the facility for parallel writes in deferred write processing.
Evaluating the analysis results
If the following condition is true, use of the facility for parallel writes in deferred write processing is considered to be effective:
  • Execution time (value of DWTOTAL + DWTOTALM) > parallel WRITE time (value of DWPARA + DWPARA)
For accurate checking, measure the execution time value of DWTOTAL + DWTOTALM) at the following times:
  • Before and after the facility for parallel writes in deferred write processing is used
  • Before and after the pd_dfw_awt_process operand value is changed
If the execution time is reduced, use of the facility for parallel writes in deferred write processing is considered to be effective.
Actions to be taken
If use of the facility is effective, perform the tuning described in 21.3.3(3) Increasing the number of parallel WRITE processes during deferred write processing.
If the use of the facility is not effective, do one of the following:
  • Stop using the facility for parallel writes in deferred write processing.
  • Restore the pd_dfw_awt_process operand to its original value

(4) WRITE unit time minimum (DWMIN DWMINM)

This is the shortest write time on a page when write operations occur on multiple pages. DWMIN displays the time in seconds, and DWMINM displays only the microseconds portion of the time.

Purpose
This information is collected in order to check the validity of disk performance.
Evaluating the analysis results
Compare with disk performance for evaluation. If you use the facility for parallel writes in deferred write processing or you change the number of parallel WRITE processes, compare disk performance before and after the facility is used, or before and after the change.
Actions to be taken
If the performance is poor compared to the disk performance, use OS functions to acquire input/output-related tuning information, and then perform tuning on the basis of that information. If the disk performance was reduced after the facility for parallel writes in deferred write processing was used or after the number of parallel WRITE processes was changed, check the parallel WRITE time. If the facility is not effective, restore the previous settings that were in use before the facility was applied or the change was made.
One of the causes of poor disk performance is disk contention. Check and, if necessary, revise the configuration of disks, RDAREAs, and tables so that input/output processing is not concentrated on any particular volume. For the design guidelines for disks, RDAREAs, and the table configurations, see the manual HiRDB Version 8 Installation and Design Guide.

(5) WRITE unit time maximum (DWMAX DWMAXM)

This is the maximum write time for a page when write operations occur on multiple pages. DWMAX displays the time in seconds, and DWMAXM displays only the microseconds portion of the time.

Purpose
This information is collected in order to check the validity of disk performance.
Evaluating the analysis results
Compare with the disk performance for evaluation. If you use the facility for parallel writes in deferred write processing or if you change the number of parallel WRITE processes, compare disk performance before and after the facility is used or before and after the change.
Actions to be taken
If the performance is poor compared to disk performance, use OS functions to acquire input/output-related tuning information, and then perform tuning on the basis of that information. If disk performance was reduced after the facility for parallel writes in deferred write processing was used or after the number of parallel WRITE processes was changed, check the parallel WRITE time. If the facility is not effective, restore the previous settings that were in use before the facility was applied or the change was made.
One of the causes of poor disk performance is disk contention. Check and, if necessary, revise the configuration of disks, RDAREAs, and tables so that input/output processing is not concentrated on any particular volume. For the design guidelines for disks, RDAREAs, and the table configurations, see the manual HiRDB Version 8 Installation and Design Guide.

(6) WRITE unit time average (DWAVG DWAVGM)

This is the average write time when write operations occur on multiple pages. DWAVG displays the time in seconds, and DWAVGM displays only the microseconds portion of the time.

Purpose
This information is collected in order to check the validity of disk performance.
Evaluating the analysis results
Compare with the disk performance for evaluation. If you use the facility for parallel writes in deferred write processing or if you change the number of parallel WRITE processes, compare disk performance before and after the facility is used or before and after the change.
Actions to be taken
If the performance is poor compared to the disk performance, use OS functions to acquire input/output-related tuning information and then perform tuning on the basis of that information. If the disk performance was reduced after the facility for parallel writes in deferred write processing was used or after the number of parallel WRITE processes was changed, check the parallel WRITE time. If the facility is not effective, restore the previous settings that were in use before the facility was applied or the change was made.
One of the causes of poor disk performance is disk contention. Check and, if necessary, revise the configuration of disks, RDAREAs, and tables so that input/output processing is not concentrated on any particular volume. For the design guidelines for disks, RDAREAs, and the table configurations, see the manual HiRDB Version 8 Installation and Design Guide.

(7) WRITE count (DWEXEC)

This is the write count for each operational cause (CAUSE).

Purpose
This information is collected to determine whether or not deferred write processing is delayed for the following reason:
  • The number of update buffers that can be processed is exceeded within the synchronization point collection interval.
Evaluating the analysis results
If the operational cause (CAUSE) is S (synchronization point), check to see if the WRITE count (DWEXEC) is equal to or less than the value obtained from the formula below. If the WRITE count is greater than the value, take appropriate action, as described below.
{synchronization point collection interval[Figure] average WRITE unit time (DWAVG, DWAVGN)} [Figure] 0.1
Obtain the synchronization point collection interval from the difference between the output times of the following messages:
  • KFPS02183-I (message displayed when synchronization point processing is completed)
  • KFPS02179-I (message displayed when synchronization point processing is skipped)
Actions to be taken
Tune by reducing the number of update buffers during synchronization point processing. For details about the tuning method, see 21.3.3(1) Reducing the deferred write trigger interval and 21.3.3(2) Increasing the update page output rate during deferred write trigger.
If the tuning does not help, specify the pd_dfw_syncpoint_skip_limit operand and limit the number of update buffers. In this case, however, if the skip count for the synchronization point processing reaches the maximum value, the update buffer is output at an extension of the update transaction, thereby adversely affecting the update transaction performance.