[Date Prev][Date Next]
Re: Benchmarks at Network World
On Mon, 22 May 2000, Julio [iso-8859-1] Sánchez Fernández wrote:
> "Booker C. Bense" wrote:
> > - Unless there has been a drastic reengineering since the Umich3.3
> > days, slapd without threads would be a single process server. It
> > wouldn't even benefit from the OS level parallelism of the subprocess
> > forking model. Basically the code used a "thread" abstraction at the
> > server level that just ran the whole thing in a single process if
> > there was no support for threads in the underlying OS. It has been
> > quite a while since I looked at that part of the source, but I
> > remember thinking at the time that it would be very difficult to
> > modify it to use the standard unix fork/exec model of server.
> Absolutely correct, that's the way it works.
> > > Any idea on what might be producing that? Besides saturation on the
> > > client side, of course, that I will presume for now it is not the
> > > case.
> > - This is exactly what I would expect without native OS threading
> > support.
> Can you explain this? Wesley Craig has suggested a way for it to happen,
> but I think it can only be seen with real, preemptive threads.
- I misread your message. I thought the time per client was scaling
linearly, not the responses per second. If the clients are all making
the same request, then the data is cached in memory (maybe, I haven't
looked at this part of the server in a long time). This might explain
the results. IMHO 10 is not really a significant number of clients to
test any server. I couldn't find details of how the tests were done,
so speculation is useless. If they are getting linear performance,
then they aren't stressing the server. To me the interesting point is
when the curve "stops" being linear, any test that doesn't give you
this point on the curve is pretty useless.
- Booker C. Bense