Disk performance issue

[Gordon_3]

Hi All,

I have a query on disk performance and hope those experts here can share point of view. I have a program which has a performance issue running on HPUX connect to Clarion disk. In glance, I can see the disk containing the DB "somtimes" has disk queue, ( <10 ) last for around 10sec. in serveral instance. So I dig out deeper in the disk level, which using Clarion disk monitoring tools. I got below result which is quite strange. In disk level, the utilization is ok, around 50% on that LUN, NO disk queue report, the number of I/O is around just 2xx something. However, I notice the count "service time" sometimes can grow up to >500 ms, so in norm. situation, service time > 500ms should be quite high, right? If yes, then is it make sense that high service time does not mean it has high disk queue and utilization as indicate in the montoring tools?
Also do u think this is the root cause for the slow? Thx for the sharing.

Gordon

[Jakes Louw · Accepted solution]

Hi Gord

In DB access, as you probably know, you have 2 types of I/O: "select" type and an "insert/update" type. The former is a "read", the latter a "write". What you have to do is determine what ratio of read to write is present when the disk service time increases. Remember that Clarion, like all arrays, uses memory as a cache, and the algorithm used is IMHO sometimes a little stupid. I have seen EMC arrays go into what they call a "priority destage". This occurs when the cache cannot be updated fast enough (read), or when the cache cannot dumped to disk fast enough (write). With today's UltraSCSI and Fiber Channel front-end bus adapters connecting into massive arrays, we sometimes forget that the backend disks are still poor old Single-ended SCSI disks running at moderate I/O rates. The server-based monitoring software then gets conflicting stats from various devices in the chain, and it takes some analysis to determine the real problem. I haven't used Calrion before, but I must assume it has it's own monitoring software? Check if you are getting cache,bus, or spindle contention on the backend.