HP-UX 11i v3 Mass Storage I/O Performance Improvements

• escsi_maxphys — The maximum I/O size allowed on the system. This tunable replaces the
scsi_maxphys system tunable available on previous releases of HP-UX. The default setting is
increased from 1 MB to 2 MB in HP-UX 11i v3, with corresponding interface driver enhancements
to support the larger size.
The Logical Volume Manager (LVM) is enhanced in HP-UX 11i v3 to support the larger I/O sizes,
as discussed in LVM improvements
.
• max_q_depth — The maximum number of I/O requests outstanding per LUN path. This replaces
the previously available scsi_max_qdepth system tunable available on previous releases of HP-
UX. The new max_q_depth SCSI tunable provides more flexibility in setting the queue depth. It
can be set globally, as before, or per device or per device type. It can also be set based on other
criteria such as vendor ID and product ID.
This flexibility is important because the performance characteristics of increasing or decreasing this
tunable are device specific. Some devices or device configurations have more internal I/O
resources than others and have varying algorithms in how their resources are used. Setting the
max_q_depth too high can result in increased CPU utilization; setting it too low can unnecessarily
limit the I/O rate. The default value is 8.
HP conducted a test using the Disk Bench I/O benchmark tool on an rx6600 server with a 2-port 4
Gb/s AB379B Fibre Channel adapter connected to MSA1500 Fibre Channel disks (2 paths per
LUN). The load factor per LUN (number of processes performing I/O to each LUN) was 32. The
results showed IOPS improvements when max_q_depth increased from 8 to 16, particularly with
reads. Increasing max_q_depth from 16 to 32 in these tests produced either no increase in
performance or a slight decrease. Table 1 shows these results.
Table 1: IOPS as max_q_depth increases
Test
Qdepth
=
4
Qdepth
=
8
Qdepth
= 1
6
Qdepth
= 3
2
1K Read IOPS, mpath off
251,000 297,000 304,000 304,000
1K Read IOPS, round-robin
294,000 299,000 302,000 302,000
1K Read IOPS, least-cmd-load
290,000 301,000 304,000 302,000
1K Write IOPS, mpath off
211,000 278,000 295,000 284,000
1K Write IOPS, round-robin
280,000 298,000 296,000 296,000
1K Write IOPS, least-cmd-load
279,000 296,000 290,000 288,000
The mpath off label in Table 1 refers to tests with native multi-pathing disabled, while round-
robin, and least-cmd-load refer to tests with multi-pathing enabled using the respective load-
balancing policies. For comparison purposes, the same set of legacy DSFs balanced across two HBA
ports was used in all these tests. The tests on 11i v3 were also performed using the corresponding
new persistent DSFs, without affecting the results.
LVM improvements
A number of LVM performance improvements have been provided in HP-UX 11i v3, including:
• Large I/O support. LVM now supports I/O sizes up to the extent size, within the limit of the
escsi_maxphys setting and the HBA driver support as discussed in New SCSI tunables
.
• Faster resynchronization via an enhanced Mirror Write Cache (MWC). The MWC has been
enhanced in HP-UX 11i v3 to support larger I/O sizes and to increase the size of the cache.
10