[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference ssag::ask_ssag

Title:	Ask the Storage Architecture Group
Notice:	Check out our web page at http://www-starch.shr.dec.com
Moderator:	SSAG::TERZAN

Created:	Wed Oct 15 1986
Last Modified:	Fri Jun 06 1997
Last Successful Update:	Fri Jun 06 1997
Number of topics:	6756
Total number of notes:	25276

6623.0. "CREO DIBENCH performance with HSZ50 and 9GB drives (RZ40)" by MQOOA::COURCHESNE (Bon renom et loyaute) Thu Apr 24 1997 10:39

    I have a bir performance problem.  I'm wondering if anybody has an
    idear.  I'm a bit at a loss here and 50 RAID Array 450 are at stake
    here.
    
    Here is the config
    
    AlphaServer 4100 - 512 MB of memory
    NT V4.0 no service pack installed
    One KZPSA
    One HSZ50 - 32MB cache
    One RAID450
    7 - 9 GB drives spreadded out on all channels
    RAID 3/5 config with all 7 drives
    
    I'm running a benchmark that does the following operation in huge
    files ( more than 400MB):
    
    first pass: 64K block reads
    second pass: 64K block Reads
    Third pass:10K block reads
    fourth pass: 64K block writes
    fifth pass: 64K block wrrites
    
    The benchmark, called CREO - DIBENCH (disk index benchmark), gives me
    back a throughput of about 4MB/s with the above config.
    
    I need to go faster.  I've tried bringing the chunksize to 16, (from
    the original 256) and the test ran longer.  I'm now trying the same
    test with a chunksize of 64 (I'll keep you posted).
    
    Any idear on what I could do to improve?
    
    Thanks
    Andre J. Courchesne
    Digital Montreal

T.R	Title	User	Personal Name	Date	Lines
6623.1	Benchmark available	MQOOA::COURCHESNE	Bon renom et loyaute	`Thu Apr 24 1997 10:41`	4
	If you want, I can provide the benchmark files and procedures. Andre J. Courchesne Digital Montreal
6623.2	Just about right, unfortunately...	EYLAK::BATES	Ken Bates	`Thu Apr 24 1997 11:20`	31
	We recently completed a customer benchmark on VMS that did essentially what your benchmark did, and got very similar results. The problems we found, which may o rmay not be applicable to yours, were: o Synchronous I/O means no queue at the disks, so the disks are idle for a fair amount of the time. o No spindle synch among the drives means substantial rotational latency, adding to the low I/O performance. o If you change to RZ28 (plain, not M or B), you should increase your performance by about 20% due to their superior read-ahead cache algorithms. o Lack of controller read-ahead cache means you are running at essentially disk speeds. Bottom line: You can get higher performance by having multiple streams access the disk (asynch I/O). Other than that, each transfer will result in some amount of rotational latency (variable, depending on the specific I/O), and some two disk transfers (chunk crossings with additional rotational latency). You probably won't get much better than you're getting right now, short of increasing the transfer size to a value greater than or equal to the number of members in the RAID set multipled by the chunk size, where the chunk size is equal to the track size. I don't know the disk types, but you might try a chunk size equal to the track size of the outer band, locate the file at a very low LBN, then issue I/O with a transfer size of 1024 sectors or greater. That should help. Then again, depending on the O/S drivers, it may not... - Ken
6623.3		OTOOA::LAVIGNE		`Thu Apr 24 1997 17:15`	14
	The other thing that hasn't been mentioned about this test is that the DG Clarion product is over twice as fast as the Raid 450. How is it that the DG Clarion controller is not affected the same way. We need the proper resources to fix this or we can kiss 2-3 million $ good-bye. My understanding is that Bob Hazlet the northern business unit storage technical support person is going after the resources in Colorado and Shrewsbury to get this fixed. Ken, if there is anything you can add please do so. Regards, JP Lavigne Storage SPecialist Eastern Canada
6623.4	7.3MB/sec and counting	MQOOA::16.174.160.12::courchesne	Bon renom et loyaut�	`Tue Apr 29 1997 14:03`	8
	Some performance enhancement has been achieved by setting a flag in the HSZ controller. A async mode has been enabled and the HSZ got performance of about 7.3MB/sec, we'll be running again with this flag set and with a different chunksize and we'll see what type of performance we'll be achieving. Andre J. Courchesne Digital Montreal
6623.5	reformatted to < 80 chars	BIGUN::KEOGH	I choose to enter this note now.	`Mon May 05 1997 01:59`	10
	Some performance enhancement has been achieved by setting a flag in the HSZ controller. A async mode has been enabled and the HSZ got performance of about 7.3MB/sec, we'll be running again with this flag set and with a different chunksize and we'll see what type of performance we'll be achieving. Andre J. Courchesne Digital Montreal