This is really only a partial summary since response was
rather sparse, but for what it's worth:
> [...background detail removed...]
> My speculation is many of the physical reads on this domain were
> satisfied by UBC even though they were not satisfied within
> Oracle's buffers.
> Can anybody offer an opinion?
Yes, satisfied by UBC.
Since Oracle "physical" reads really means reads passed to
the operating system, tuning of UBC can affect Oracle performance.
> *3 Does anybody know why hsz dstat failed to report the
> write command rates?
No response on this.
This workload was primarily read. When running dstat on a more active
system the write command rates were reported.
Since dstat is unofficial, particulary for hsz40, not worth
pursuing further.
> *4 For the two active domains on the monitored hsz40's, the HSZ
> read command rate significantly exceeded the uaiostat transfer rate.
> During many of the intensive 5 minute reporting intervals
> the ratio was close to exactly 1.50 hsz read commands to
> UNIX transfers. Likewise, if you subtract out the Oracle write
> from the uaiostat transfers the ratio also approaches 1.50 for
> the entire period.
> My only guess is the hsz40 'read-ahead' algorithm may have been
> invoked on these domains since the tests primarily reflect
> serial reads through large data files.
> Can anybody offer an opinion?
No response on this.
> *5 Does anybody know where there is a good discussion
> (short of reading the UNIX source code) on the vpf_ubc* fields?
No response on this.
We have started reading the source code to augment the manuals.
By appearance some of the other tuning parameters we've previously
left alone merit further examination:
including ubc_maxdirytwrites, vm_ubcbuffers, vm_syncswapbuffers,
vm_asyncswapbuffers, and particularly vm_ubcseqstartpercent.
We're starting to look at this but first working on a monitoring
tool to help gauge how effective tweaking these will be.
If anybody has some examples and discussion on adjusting these,
I'd be interested in hearing your experience.
> *6 The values reported by vmubc were meaningless,
> Table#3 came from vm_perfsum.
> Can anybody offer an opinion on vmubc and UBC usage analysis,
> specifically as it relates to AdvFS?
AdvFS does utilize UBC somewhat differently than UFS
(from a colleague reading the source).
> #7 Does anybody know of any good reference material on breaking
> out IO performance components (Oracle, UNIX, AdvFS, UBC, kzpsa,
> hsz40, devices) along the entire path?
This question was previously asked by another, he got no response.
> Does anybody know of good io benchmarking materials for
> Oracle under Digital UNIX?
> (I would be happy to sign non-disclosures if necessary).
No response on this.
On a side note, while waiting for installupdates running for
v3.2g->v4.0b last weekend I passed the time by rewriting iostat
from scratch. Besides the enhancements we had previously added
as uaiostat for hsz support and summaries, I've added
a horizontal display since we now have systems which cannot
show all their disks in 132 columns. If anybody else has some
pet gripes with iostat which they'd like tucked into a replacement
utility, let me know and I'll see if I can work them in.
Basic command syntax is compatable with iostat.
A sample:
sxkac_at_glacier> uaio -St -h6 60 4
Boot: 97/04/20 97/04/30 | Seconds : 60 60 60 60
10:05:58 11:14:14 | 240 | 11:15:14 11:16:14 11:17:14 11:18:14
glacier transfer tps transfer tps transfer transfer transfer transfer
Bus# 0: 1: . . . . . . . .
Bus# 3: 4: 17271465 20 19823 83 4897 5299 5196 4431
Bus# 6: 4: 1066812 1 . . . . . .
Bus# 7: 5: 16297545 19 17559 73 6223 1875 1947 7514
Total 4: 14: 34635822 40 37382 156 11120 7174 7143 11945
Disk rz0 : . . . . . . . .
Disk rz25 : 5575342 6 5546 23 1422 1456 1393 1275
Disk rz26 : 3178761 4 1727 7 1265 179 167 116
Disk rz27 : 6515573 8 10701 45 1798 3329 3208 2366
Disk rz28 : 2001789 2 1849 8 412 335 428 674
Disk rz48 : 305967 0 . . . . . .
Disk rz49 : 294580 0 . . . . . .
Disk rz50 : 290914 0 . . . . . .
Disk rz51 : 175351 0 . . . . . .
Disk rzb57 : 3464076 4 5278 22 2262 346 61 2609
Disk rzc57 : 4320239 5 4164 17 787 548 1434 1395
Disk rz58 : 2709392 3 1617 7 671 117 99 730
Disk rz59 : 2726999 3 3518 15 1304 699 152 1363
Disk rz60 : 3076839 4 2982 12 1199 165 201 1417
cpu User 410028.4 12% 477.3 50% 117.5 109.7 125.7 124.4
cpu Nice 17157.0 0% 2.1 0% 0.1 0.0 1.9 0.1
cpu System 330583.1 10% 175.6 18% 40.9 39.3 45.9 49.5
cpu Wait 630270.0 18% 0.0 0% 0.0 0.0 0.0 0.0
cpu Idle 2084341.3 60% 305.1 32% 81.6 90.9 66.6 66.0
I'll make uaio available for ftp when I'm done testing changes to summarize
previously archived runs of the program.
_____________________________________________________________________
Kurt Carlson, University of Alaska SOIS/TS, (907)474-6266
sxkac_at_alaska.edu 910 Yukon Drive #105.63, Fairbanks, AK 99775-6200
Received on Wed Apr 30 1997 - 22:12:14 NZST