2. PERFORMANCE REPORTING ANALYSIS › 2.4 I/O Configuration Analysis › 2.4.4 DASD I/O Component Analysis Inquiry › 2.4.4.2 Technique Tutorial › 2.4.4.2.2 I/O Component Analysis for MVS/370
2.4.4.2.2 I/O Component Analysis for MVS/370
DASD I/O response time is the sum of a number of unique
events, as is shown in the diagram in Figure 2-98. The
actual events are shown above the time line, and the
calculated components of I/O response time are shown below.
(Note that the asterisks on the time line indicated data
transfer.) The first technique, published by IBM for MVS/370
systems, breaks I/O response time into four distinct
components:
o Queue - The sum of the time IOS spent queuing for the
device and the logical channel.
o Control Unit - The time that the channel was available,
but the control unit(s) for the device were busy.
o Device - The delay to reach the device caused by
reserves from other systems or head-of-string
contention.
o SIO - The time that the I/O was active on the device.
This value includes seek, set sector, latency, RPS
reconnect, and data transfer.
The values for each of these four components can be
approximated from standard RMF measurements.
I/O Response Time Components
Wait for other users
| Wait for Channel
| |Wait for Control Unit
| ||Wait for Head of String
| |||Wait for Device
| |||| Seek CCW
| |||| | Seeking
| |||| | | Set Sector CCW
| |||| | | | Latency
| |||| | | | | RPS
| |||| | | | | |Reconnect
| |||| | | | | |
| |||| | | | | | XFER
| |||| | | | | | |
time line |================*========*========*****|
| || | |
|Queue||Device| SIO |
|<--->||<---->|<----------------------->|
| || |
| ->||<- Control Unit |
| |
| |
| Total |
|<------------------------------------->|
Figure 2-98. I/O Response Time Components
The four components of I/O response time are computed from
six standard CA MICS data elements:
o DVAPCACT - Device percent busy
o DVAPCRDL - Device percent reserve delay
o DVAPCCUB - Percent control unit busy
o DVAAVQUE - Device average queue length
o DVASIO - The total number of SIOs serviced by the
device
o DURATION - The length of the interval
For additional information on these values, consult the
CA MICS Data Dictionary.
The total number of start I/Os and the duration are used to
compute the start I/O rate:
SIORATE = DVASIO / DURATION (Equation 1)
The SIORATE value will be used in equations in the following
paragraphs.
SIO time is computed based on the percent device busy value
and the start I/O rate. This calculation is shown below:
SIO = (DVAPCACT/100) / SIORATE (Equation 2)
The most significant portion of SIO time is device rotational
delay. For devices on busy paths or strings, the path to the
CPU can be busy when the device attempts to reconnect to the
channel. This causes the reconnection to fail, and the I/O
must then wait a complete rotation (16.7 milliseconds for
most device types) before another reconnect can be attempted.
RPS reconnect probabilities are discussed in the DASD skew
technique tutorial (Section 2.4.5.2).
Device time is computed based on the percent reserve delay
and the start I/O rate. This calculation is shown below:
DEVICE = (DVAPCRDL/100) / SIORATE (Equation 3)
Device time, that is,the percent of time the system could not
access the device even though the UCB (unit control block)
indicated that it was free and the channel path and control
unit were free, is caused by three problems:
o Shared DASD contention
o Reserves issued by another system
o Head-of-string contention
The first two of these problems can be investigated using the
Shared DASD Analysis Report. The DASD Skew Analysis Report
can provide insight into head-of-string contention problems.
Control unit delay occurs when the channel is free, but the
control unit is busy supporting an I/O to another device or
string. Control unit delay is calculated based on percent
control unit delay and the start I/O rate. This calculation
is shown below:
CNTLUNIT = (DVAPCCUB/100) / SIORATE (Equation 4)
In most cases, this time is very small, that is, less than
five milliseconds.
Unfortunately, the internal queuing time cannot be directly
estimated. However, we can exploit the fact that we have
already computed the other three components needed to derive
total response time. The total response time is computed
based on the device busy value, the average queue length, and
the start I/O rate. This calculation is shown below:
TOTAL = (SIO + DVAAVQUE) / SIORATE (Equation 5)
Knowing the total time, we can then calculate the internal
queuing delay, as shown below:
QUEUE = TOTAL - DEVICE - CNTLUNIT - SIO (Equation 6)
High values of internal queuing are caused by two problems:
o Very high I/O request rates to the device. To solve
this problem, you should migrate data sets to other
devices to balance I/O activity. Most often, this
problem results from an application putting multiple
high activity data sets on a single device. You can
identify the performance group using a device by
investigating the EXCP counts in the BATWDA File.
o Significant queues for the logical channel. You can
determine this problem by investigating the average
queue length value of the standard RMF logical channel
report that may be produced using CA MICS' standard RMF
reports compendium program. If this value is high, you
may have to reconfigure your logical and physical
channels.
