iostat and maximum I/O throughput test
Hi all,
We are testing the I/O performance on a V490 conected to a SE 3310 SCSI storage array as we are experiencing a slow I/O. The test used the format command with analyze -> read option on the disk, and by watching the iostat we got around 70 MBytes/Sec of read performance, knowing that the SE 3310 SCSI is an Ultra 160 SCSI, meaning that it should go around 150 MBytes/Sec , the question here is this avalid I/O test or is there another way of testing random and sequential read/write over a physical disk or a file system and measure it? we do not want to rely on an FTP session or file copy to determine the maximum I/O throuput , we need to conduct a more specific measure. Any ideas?
[709 byte] By [
goSolarisa] at [2007-11-26 15:59:53]
# 2
Thanks Brian
Actually the disks inside the SE 3310 are also Ultra 160 SCSI disks so both the HBA and the disks are of the same SCSI level, I know the 160 MB/s is thoritical but I am expecting to get 90% of this I/O limit. It could be that the I/O test I am conducting is not the proper one, I am looking for a real I/O performance measue to help in making sure there is no problem with the setp.
# 3
Why would you be expecting to get 90% of the bus throughput.
The bus throughput is about the last thing to look at for expected throughput.
Your usually limited by the rate the platter passes by the read head
Start with the multiplying the rotational speed by the average track capacity.
But thats only a single disk. You should be worrying about the aggregate throughput from
multiple disks.
How are you aggregating the disks. software raid0, hardware raid 5?
You can use dd to the raw disk to get data throughput. I prefer to do write testing
copy a chunk of /dev/zero to the disk to avoid caching issues.
Then do the same with a virtual aggregate disk and check how the throughput ramps up.
# 5
Well, if you talking about single disk thoughput. Most modern disks can manage a sustained throughput of about 50-60 Megs per second. Thats to the raw disk.
And probably a bit less once you add filesystem overhead.So if your getting that then your successfully saturating the disk.
Thats generally not hard. So what you should be worried about is getting suitable aggregate throughput once you stripe the disks and add mirroring or raid.
