home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!poietics!cng
- From: cng@poietics.UUCP (Chris Gayler)
- Newsgroups: biz.sco.general
- Subject: Re: Performance issues revisited
- Message-ID: <460@poietics.UUCP>
- Date: 23 Dec 92 13:10:26 GMT
- References: <sheldon.725038896@pv141b.vincent.iastate.edu> <Duw8VB10w165w@zswamp.UUCP>
- Reply-To: cng@poietics.UUCP (Chris Gayler)
- Organization: POIETICS
- Lines: 56
-
- In article <Duw8VB10w165w@zswamp.UUCP> geoff@zswamp.UUCP (Geoffrey Welsh) writes:
- >sheldon@iastate.edu (Steve Sheldon) writes:
- >
- >> time dd if=/dev/rhd0a of=/dev/null bs=131072 count=128
- >
- > I know that the DMA controllers on PC-compatibles can't do a transfer of
- >more than 64K in a go. In theory, the Adaptec (models 154x, 164x, 174x) need
- >not be limited by this, thanks to bus master DMA. Does anyone know the size
- >limit for a single DMA burst from these devices?
- >
- >> I'm curious if perhaps the configuration you have just doesn't like the
- >> smaller block size. I suppose it might be possible...
- >
- > In my own experiments while constructing a count-oriented (for maximum
- >performance) yet time-regulated disk benchmark, I found that throughput
- >increases logarithmically with block size. (note: not exponentially; I mean
- >approaching the maximum, with each increment in block size having less effect
- >than the last.) I ended up using 60K blocks, because they gave pretty much
- >the same figures as 63K blocks. Some systems gave very low figures when 64K
- >block sizes are used, but others did not; I do not understand that
- >observation.
- >
-
-
- Just to add more datapoints, I ran these tests against various controllers:
-
- time dd if=/dev/rhd0a of=/dev/null bs=2048 count=8196
- time dd if=/dev/rhd0a of=/dev/null bs=65536 count=256
-
- Results (real times in seconds)
- 2048 65536 (Block sizes)
-
- 20 -- 486/50 with Mylex DCE376 SCSI caching controller, SCO 3.2.4
- Newer Maxtor disks.
-
- 33 10.6 486/50 with Adaptec 1740 SCSI, SCO 3.2.4, older Maxtors
-
- 34 33.9 386/25 with WD1007 ESDI controller, SCO 3.2.2
-
- 37 10.1 486/33 with Adaptec 1542 SCSI, fast Fujitsu disks, SCO 3.2.4
-
- 39 23.9 386/25 with UltraStor 12F ESDI controller, SCO 3.2.2
-
- 39 13.3 386/33 with Adaptec 1542 SCSI, fast Fujitsu disks, SCO 3.2.4
-
-
- To summarize, SCSI and ESDI controllers have comparable performance on small
- block sizes, but SCSI steps way out on larger block sizes. The exception is
- the Mylex, which likely has the best performance on both ends (I didn't have
- access to the Mylex to run the large block test.) For the 1542, the mem=/L
- boot option yielded about an 8% improvement on the 2048 test.
-
- The downside is that the Mylex will bite you on installation time. :-E
-
- Chris
-
-