>-----Original Message----- >From: Sander Temme [mailto:[EMAIL PROTECTED] [SNIP] > >> There's one more thing I noticed (might be specific to >HP-UX) : I saw more >> errors with keepalive ON rather than when it was OFF. > >I think you may be looking at a problem where server and client(s) are >messing with each other. Threads blocking on the server could >be tying up >the clients, untill all clients are hanging and everything >stops. I have >found the SPECWeb99 clients less than wonderfully robust under these >circumstances.
I don't know - with keepalive turned ON, I see the server is hovering around 50% idle - but still not giving the enough throughput.. OTOH, the client(s) is 70% idle. I saw that server has a whole bunch of TIMED_WAIT socket states (around 2000), but the client(s) had only a few of them. I thought the server was running out of sockets (or something really bad happening), and hence turned the keepalive OFF. Another thing that I started noticing : if I enable many threads (1000 threads) per process (on 1-way box), the performance is worse than enabling fewer threads (100 threads) per process. There were far too many semops and semtimeouts. I ended up having 7000 - 8000 context switches !! -Madhu