I/O performance problem when reading.

[Yves Sevigny]

Hello,

I just installed a BA356 disk enclosure with 7 x 18.9 GB JBODs to my Alphaserver DS20 (Tru64 V5.1B, scsi controller 3X-KZPCA-AA).

When I wrote a 9GB file to one of the disks, it took around 4 seconds. But, when I copied that 9GB from that JBOD in the BA356, it took 54 seconds!! Did the test on 4 of the disks with the same result.

I did some disklabel -r on those disks and the only thing different I saw was that some of the disks had the flag FLGS: Dynamic-geometry.
Did some research but did not find if it could be the cause.

Any clue why the reading is so slow?

Thanks,
Yves

[Ivan Ferreira]

If you want to do performance tests, you should try with raw devices.

I don't know if I understood whell, but when you write 9 GB to the disk is fast and when you read is slow? This may be becuase of the file system cache. Writes can go to the UBC and then flushed to disk.

[Yves Sevigny]

Hello Ivan,

It is correct, writing is fast and reading is very slow.

I apologies, I wrote 9GB file, it is a 980MB file.

For the file system cache, where do I check for this?

Yves

[Ivan Ferreira]

You can use collect -sm to view the UBC usage. Also, you can compare with the data written to disk using collect -sd, or combine both using collect -sdm.

[Yves Sevigny]

Thanks for the collect command.

During the tests, the UBC fluctuated between 71 and 132. Is this good or bad??

The transfer rate for reading is 1.7MB/sec
and for writing is 3.5MB/sec. The disks busy
rate during the xfer was 99.85%.

My log files are error free.
I was wondering if there is a kernel parameter or something that needs to be tweaked?

[Hein van den Heuvel]

There is nothing to be tweaked or tuned other then your understanding and expectation.
Your system is just fine living up to the specifications

>> The transfer rate for reading is 1.7MB/sec
and for writing is 3.5MB/sec. The disks busy

Why do you say that?
If you manage to read 980 MB in 54 seconds then that translates to 18MB/sec. This is a fine rate for a normal, older, single disk on a SCSI 2 connection (max = 20MB)

How did you create the data to write from?
Program generated? Anyway, you did NOT measure the time to write the data to disk, you measured the time to release that data to the cache. The IOs woudl have been ongoiing for almost 60 seconds after teh program finish.
A fine write-to-cache speed it was, but it would have dropped to 18MB/sec (or a little less) if you had trully written 9GB, exceeding your cache/physical memory.
Or... if you had followed the program with a a 'sync' command and waited for that to finish.

Cheers,
Hein.