summy of some 9.1.0 perf work...
Rick Jones
raj at cup.hp.com
Fri Feb 9 00:29:15 UTC 2001
Here is a summary of some of the perf measurements I've made thusfar on
the 9.1.0 named. This follows a similar methodology to the briefs
already published at:
ftp://ftp.cup.hp.com/dist/networking/briefs/
Here is the summary chart from one of the briefs there:
8.1.2 8.2.2pl5 8.2.2pl5 8.2.2pl5
"HP" "stock" "-O" "full"
System +====================================================
L2000 1x440 | 9,863 | 3,529 | 5,935 | 8,772 |
L2000 2x440 | 10,984 | 3,685 | 6,309 | 9,507 |
J5000 2x440 | 10,550 | 3,670 | 6,421 | 9,436 |
J5600 2x552 | 13,898 | 4,598 | 8,002 | 12,032 |
+====================================================
and here is the 9.1.0 stuff.
BIND 9.1.0 named Performance Summary HP-UX 11.ACE
netperf DNS_RR test requesting 1000 out of cup.hp.com
+O2 +O2
"stock" "+O2" +DD64 nothread "full"
System +=================================================+
L3000 1x550 | 2,065 | 3,127 | 3,039 | 4,303 | |
L3000 2x550 | | 5,338 | | | 6,771 |
L3000 4x550 | | 8,065 | | 4,757 | |
+=================================================+
I hope to take some quick measurements on a J5600 to have an
apples-apples comparison.
I used the HP ANSI C compiler.
"stock" is as the bits compile out of the tarball - -g, not much else.
"+O2" is replacing the -g with +O2 (aka -O)
"+O2 +DD64" is a 64-bit compilation with +O2
"nothread" is with +O2 and configure --disable-threads
"full" in the 9.1.0 context is +O4, PBO (profile-based optimization),
static branch prediction, and large page sizes for text and data
segments. in the context of 8.2.2pl5 it was also a 64-bit compilation.
I expect at some point to be able to kick-loose the entire write-up
along with some profiles and processor performance counter data. those
used hp-internal tools so I'm just trying to get the i's crossed and t's
dotted first :) when I do, it will probably be at the url
ftp://ftp.cup.hp.com/dist/networking/briefs/bind9_perf.txt. I can run
the whole document past bind-workers first as a sanity check before
going public (unless folks think it would be wrong to send the whole
thing to bind-workers...)
The user/kernel split was roughly 80/20. One thing I did notice was that
the number of select calls appears to match the number of transactions.
I'm guessing (without first looking at source :) that the tune from
8.2.2 to do multiple recvfrom's per select is not in the 9.1.0 bits -
not sure if it applies, but it might...
comments/suggestions welcome,
rick jones
--
ftp://ftp.cup.hp.com/dist/networking/misc/rachel/
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to email, OR post, but please do NOT do BOTH...
my email address is raj in the cup.hp.com domain...
More information about the bind-workers
mailing list