Jeffrey Burgoyne wrote:
> I'm not really sure why the big difference. Aound January our performance
> with 1.2.6 was really dropping to the point everyhing was unusable. We
> then added a second CPU to the DEC box, and got very interimitent
> performance. It would go along great for a few minutes, then slow down to
> a crawl. At times during the night (when there was a lot of processor time
> available), the server would sometimes take ten minutes to serve up a
> page.
>
> About the best I could figure was that the SSL or Apache's multi-threading
> had some problem.
I've seen some similar random symptoms and weird problems, but with
Apache 1.3.6 and recent versions mod_ssl. I can't pin it
down either. For example, just yesterday a server appeared
"hung" and I went poking around with truss, lsof, netstat, etc.
There were only 5 children, and 3 were blocked on reads (keepalives I
assume), one blocked on a poll, and one on fcntl. But no
child was on 'accept', which accounts for the server being "hung"
of course. After a while, but definitely more than
my 15 second keepalives timeout, things "clear up", i.e.
some child gets free and goes back to accept'ing. The machine
itself is essentially idle and load appears to have little to do with it.
If anyone has any clues, I'd be delighted to hear. I frankly
don't even know where to look.
But this is a tangent to my observation about the big differences
in performance estimates, i.e. Victor says 10-20 conn/sec
while you say you see 500 conn/sec. Obviously, conn/sec is highly
dependent upon machine. Victor didn't qualify what machine he
meant 10-20 would be the max, but I assume he meant something
like a typical single processor Linux x86 system (400-500Mz).
If you're getting 25-50 times that on a two processor alpha,
something other than the machine would have to account for
the difference, no?
Mark
> After upgrading the intermitent dropouts went completely away. Loads are
> well distributed across the CPU's. What caused the dropouts was never
> discovered, but it made it very hard to guage the previous performance
> accurately. By looking at the best response times, we still had a
> performance gain, but more in the magnatiude of 50%.
>
> Jeff
>
>
> On Wed, 4 Aug 1999, Mark Dedlow wrote:
>
> >
> > Victor Khimenko wrote:
> >
> > > If you need SSL speed is not issue anymore. SSL is VERY processor-intensi
ve
> > > so you'll got at most 10-20 connections per second. Additional timeout fr
om
> > > ping-pong between 1.2.6 and 1.3.6 will be dwarfed by SSL timeout on any
> > > decent OS...
> >
> >
> > However, last week someone on this list said:
> >
> > > I must say that the performance boost we got when moving from stronghold
> > > to mod_ssl (1.2.5 apache to 1.3.6) was quite signifigant. At peak we
> > > handle over 500 connections per second with no performance degredation
> > > (Two CPU Alpha).
> >
> >
> > How to account for the humongous difference?
> >
> > Mark
> > ______________________________________________________________________
> > Apache Interface to OpenSSL (mod_ssl) www.modssl.org
> > User Support Mailing List [EMAIL PROTECTED]
> > Automated List Manager [EMAIL PROTECTED]
> >
>
> ______________________________________________________________________
> Apache Interface to OpenSSL (mod_ssl) www.modssl.org
> User Support Mailing List [EMAIL PROTECTED]
> Automated List Manager [EMAIL PROTECTED]
>
______________________________________________________________________
Apache Interface to OpenSSL (mod_ssl) www.modssl.org
User Support Mailing List [EMAIL PROTECTED]
Automated List Manager [EMAIL PROTECTED]
