 |
|
A set of Linux memory scaling tests showed the following
results:
The throughput of the z/VM guest with default Linux DASD driver
drops off with 8GB memory size and more (i.e. the numbers are so
small that they don't appear in the chart).
The z/VM guest tests with the enabled fixed I/O buffer feature
show a continuing increase of throughput, similar to the behavior
of the Linux in LPAR tests.
Switching on the fixed I/O feature when running in LPAR can lead
to slight performance degradations and therefore should not be
considered.
- IBM eServer zSeries 900 (2064-216)
- IBM TotalStorage Enterprise Storage Server 2105-800
- LPAR with 8 CPUs, 14GB storage, 2GB expanded, z/VM
5.1
- Linux 64bit SLES9 kernel 2.6.5
- Linux guest 4 CPUs, 2 – 10GB memory
- Informix 9.4.0
- Benchmark Informix OLTP
- File system ext2
|
|
|
The tests showed that the throughput for a 256MB guests with
fixed I/O buffer feature switched on and off are very similar, but
we observed an increase in system CPU consumption of up to 60% when
switched on.
For a 4GB guest the throughput of the test with fixed I/O
buffer feature on should be similar to the throughput of the 256MB
guest tests because the total available Linux memory can't be used
for buffer cache. This was achieved for sequential writes and
reads. However the CPU consumption was much higher, e.g. for the 32
disk sequential write test guest CPU time increased by 75 % and CP
CPU time by 400 %. The responsible z/VM paging activity for
sequential IOZONE workloads apparently cannot be prevented by our
fix with.
- IBM eServer zSeries 990 (2084-316)
- IBM TotalStorage Enterprise Storage Server 2105-F20
- LPAR with 16 CPUs, 12GB storage, 2GB expanded, z/VM
5.1
- Linux 64bit SLES9 kernel 2.6.5
- Linux guest 2 CPUs, 256MB, 4GB
- Benchmark Iozone 1.1, write and read, file size
700MB
|
|
|
|
|
|
