HP NonStop G-Series - Manual (Page 154)

Balancing and Tuning a System
Measure User’s Guide — 520560-003
7- 18
Checking and Tuning Problem Areas
Balancing Disk Queues
Disk queues indicate the amount of time an I/O request spent waiting within the disk
subsystem. Because a CPU frequently waits for an I/O operation, disk queues can
cause major performance problems. By balancing the disk queues across all disks, the
average I/O request wait time becomes as short as possible. Because an I/O request
travels through the disk subsystem serially, calculate the total disk queue time by
adding these counter values:
•
The RECV-QTIME counter for the disk process (PROCESS entity), which
represents how long the I/O request waited for the disk process (the external
queue for the disk process). Often, a high RECV-QTIME value indicates that the
disk process resides on an overly busy CPU.
•
The REQUEST-QTIME counter for the disk (DISC entity), which represents how
long the I/O process waited for the disk (the internal queue for the disk process).
Often, a high REQUEST-QTIME counter indicates an overly busy disk.
To balance disk queues, first examine the queue length of each disk. (In Enform, you
can define a new field for disk queue lengths and use that field for balancing the
queues. In MEASCOM, you must do the calculations by hand.) When you know which
disks have long queues and which have shorter queues, balance the queue lengths by
moving files from one disk to another.
Balancing queue lengths is difficult because you must balance both disk processes
and disk queues simultaneously. In addition, you might have to work around
application restrictions and shortcomings; for example, a server program that uses
hard-coded file names instead of ASSIGN messages or DEFINEs to determine target
files to open.
In deciding which file to move and its new location, consider individual queues rather
than total queue length:
•
If the summation of the RECV-QTIME counters for all disk processes in the CPU
you are relieving is high, the CPU is overly busy. Move frequently accessed files off
the disk. Choose a file with a large number of logical I/O operations by examining
the READS and WRITES counters of the FILE reports. Move the chosen file to a
CPU with a relatively low RECV-QTIME counter; that is, a CPU with a relatively
light CPU load.
•
If the summation of the REQUEST-QTIME counters for disk processes in the CPU
you are relieving is high, the disks are overly busy. Move a file that causes
substantial disk I/O to another disk. Choose a file with a large number of physical
I/O operations by examining the DRIVER-INPUT-CALLS and
DRIVER-OUTPUT-CALLS counters of the DISCOPEN reports. Move the chosen
file to a CPU with a relatively low REQUEST-QTIME counter; that is, a CPU with a
relatively light physical I/O load.
Also consider the ease with which you can move the file. For example, alternate key
and swap files are relatively easy to move; TMF-audited files are not.