|
Contents | Previous | Next
Performance Considerations
These items warrant consideration since they have potential for a
negative impact to performance.
- Preferred Guest Support
- 64-bit Support
- FBA-emulated SCSI DASD CPU Usage
- TCP/IP VM IPv6 Performance
Starting with z/VM 5.1.0, z/VM no longer supports V=R and V=F guests.
Accordingly, if you currently run with preferred guests and will be
migrating to z/VM 5.1.0, you will need to estimate and plan for a likely
increase in processor requirements as those preferred guests become V=V
guests as part of the migration. Refer to
Preferred Guest Migration Considerations
for assistance and background information.
z/VM 4.4.0 and earlier releases provided both 31-bit and 64-bit versions
of CP. Starting with z/VM 5.1.0, only the 64-bit build is provided. This
is not expected to result in any significant adverse performance
effects because performance measurements have indicated that both
builds have similar performance characteristics.
It is important to bear in mind that much of the code in the 64-bit
build still runs in 31-bit mode and therefore requires that the
data it uses to reside below the 2G line. This is usually not a
problem. However, on very large systems this can result in degraded
performance due to a high rate of pages being moved below 2G. For
further background information, how to tell if this is a problem,
and tuning suggestions, see
Understanding Use of Memory below 2 GB
.
The FBA-emulation SCSI support provided by z/VM 5.1.0 is much less
efficient than either dedicated Linux SCSI I/O or traditional ECKD DASD
I/O. For example: CP CPU time required to do paging I/O to
FBA-emulated SCSI devices is about 19-fold higher than the CP CPU time
required to do paging I/O to ECKD devices. As another example, CP CPU
time to do Linux file I/O using VM's FBA-emulation SCSI support is
about ten-fold higher than doing the same I/O to SCSI devices that are
dedicated to the Linux guest, while total CPU time is about twice as
high. These impacts can be reduced in cases (such
as in the second example) where minidisk caching can be used to reduce
the number of real DASD I/Os that need to be issued. These performance
effects should be taken into account when deciding appropriate use of
the FBA-emulation SCSI support. See
Emulated FBA on SCSI
for measurement
results and further discussion.
Measurement results indicate that there are some cases where
performance can be degraded when communicating between TCP/IP VM stack
virtual machines using Internet Protocol Version 6 (IPv6), or using
IPv4 over IPv6-capable devices, as compared to IPv4. The most
unfavorable cases were observed for bulk data transfer across a Gigabit
Ethernet connection using an MTU size of 1492. For those cases,
throughput decreased by 10% to 25% while CPU usage increased by 10% to
40%. For VM Guest LAN (QDIO simulation), throughput and CPU usage were
within 3% of IPv4 for all measured cases. See
Internet Protocol Version 6 Support
for
measurement results.
Contents | Previous | Next
|
