Why is read-performance so slow on EVA4000 from RHEL4?

[Jesper Sivertsen]

Hello all.

I have 2 EVA4000 with 40 disks each ( 20x300Gb and 20x146Gb ) all 10K fiber-disks.

Both have two disk groups. One with 20x300Gb and one with 20x146Gb.

The 2 EVAs are located at 2 different sites, but is connected and used for HA on our ESX environment.

Firmware on both EVA4000 systems is: CR0EB0xc3p-6100.

It is mainly used for Oracle-databases and VMware ESX-Server.

All Oracle database-servers in the production environment are HP DL585's. Test environment runs on DL385. They are all equipped with HP FC2242SR 4Gb HBA ( Emulex ) for fiber connectivity.

OS is RedHat EL 4.

The BIOS on both DL385 and the HBA's are all up to date, the same goes for drivers and other software involved.

In general I can only read about 40Mb/s and write 120 Mb/s.

I am measuring with linux dd. I use a 10Gb file, so I'm sure not to hit the read/write cache.

I have tried different tricks and tweaks in linux but noting helps. ( Readahead cache, elevator=none )

I have even tried to make a single lun on each EVA and stripe them in linux, but I get the same result. This tells me that the bottleneck may not be the EVA, so it could be the san-switch, the HBA, the HBA driver, etc.

I have read a lot of threads on the internet about this issue, but I have never found a solution.

I have even had a support-case with HP, where they checked all settings and hardware and everything was ok.

Does anyone have a successful story regarding read performance from EVA4000? Where could the bottleneck be, and what is causing it?

Regards Jesper.

[Uwe Zessin]

Answer your other thread:

http://forums.itrc.hp.com/service/forums/questionanswer.do?threadId=1321211

[Ivan Ferreira]

>>> In general I can only read about 40Mb/s and write 120 Mb/s.

>>> I am measuring with linux dd.

¿Are you running your test to the raw device? To do performance tests, you should avoid doing that on file systems, do directly from the raw device.

[Steven E. Protter]

Shalom,

dd is good for making sure you can read. Its not a good performance tool.

I would look at the application side here.

A common cause of performance problems is doing things like putting a write intensive database on raid 5 LUNS. This makes writes very difficult to do and makes them take much longer.

SEP

[Jesper Sivertsen]

Thanks for your comments.
I hawe tryed both raw, fs, many luns, 1 lun.
I hawe tryed all togeter with HP, but 40Mb/s is the figure!

Jesper

[William B. Parkinson]

Try running IOSTAT in linux. Use Tlviz or csvpng to see what is going on. You could have the writes queued so high that the reads are no getting serviced quick enough for you liking. If the reads ae not cached the queue for read and write are probably first in first out. So your reads sit behind the writes and you wait. Response time, queue depth and utilization should all be looked at with IOSTAT using tlviz or csvong for linux.

[Sergio lopez_1]

Hi Jesper,
I´ve the same problem with EVA4100 (43x146GB FC 10K), BLade BL465 (Qlogic) Linux 4 and 5.2 version. Do you resolv the low performance?

[root@cppapp linux]# ./hpcreatedata -b 99614720 -n 53 -v -t -w /uno
...........................117.09 MB/sec
[root@cppapp linux]# ./hpreaddata -b 99614720 -v -d /uno
/uno/dfile000050
/uno/dfile000051
/uno/dfile000052
5035.00 MB read (53 files) in 304.06 seconds, 16.56 MB/sec

XCS 6.110 and Vraid5 (protection 1)