EVA6000 performance

[Victor Voronin]

Hi.

I can't understand how EVA works with Windows.
I try to test EVA reading speed from Windows 2003 server. I start copy to nul process, and see 35 Mb/sec speed. EVAperf as shows about 40 Mb/sec. But when I start one more copy process it shows about 30 Mb/sec speed. EVAperf as shows 80 Mb/sec.
Why one process not use 70-80 Mb/sec?

[Víctor Cespón]

Performance depends on a lot of things:

- Type of disks (FATA / 10K / 15K)
- Number of disks
- RAID level (0 / 1 / 5)
- Size of the I/Os
- Secuential or random access
- Oustanding I/Os allowed by HBA driver
- Multipath software

1) A sequential read from a sigle server and typically through a single path is not a good measure of EVA performance. This system is designed to handle a lot of reads and writes from several servers. In a sequetial read test you're not going to get more speed than from a MSA or a JBOD with the same number of disks.
2) Trust EVAperf, it knows what's going on inside the EVA. Look at the latencies and amount of I/Os on the disk group
3) Changing parameters on the HBA driver, like maximum queue depth can have a big impact on maximum throughput.

[Victor Voronin]

OK. Lets discuss

Type of disks - 15K
Number of disks - maximum possible (112)
RAID level - this volume is RAID5
Size of the I/Os - whats this&
Secuential or random access - copy to nul
Oustanding I/Os allowed by HBA driver - ?
Multipath software - HPDSM with ALB

1) A sequential read from a sigle server and typically through a single path is not a good measure of EVA performance. This system is designed to handle a lot of reads and writes from several servers. In a sequetial read test you're not going to get more speed than from a MSA or a JBOD with the same number of disks.

HP-UX server with same volume makes 100 Mb/sec/ I try to understand why?

2) Trust EVAperf, it knows what's going on inside the EVA. Look at the latencies and amount of I/Os on the disk group
On evaperf all reports read latency not more then 5 ms.

3) Changing parameters on the HBA driver, like maximum queue depth can have a big impact on maximum throughput.
Max queue depth in evaperf reports is 2. It is not more then default 32.

I dont't understand one process reading speed limits

[Víctor Cespón]

112 disks in RAID 5, and only reading data can give you a very high speed. 18000 random operations per second and 720 MB/s.

The fact that you get different results on different operating systems, while the latencies on the EVA are below 5 ms, points to a limitation on the HBA driver. The EVA is sending all the data the HBA demands, if it's not going faster is because the server is not requesting more.

I insist in that sequential read performance from a single server is not a good measure, that's not the typical usage of a shared SAN storage. You'll get the same numbers from any array with the same number of disks.