Hi Filip, that's one of the not so nice things with linux. As far as I know it's not configurable with standard linux. There exist kernel patches for this and there is an ip filter module that lets you do that, but some say that module is very bad for IP performance (and high performance would be the major reason to decrease the time_wait interval).
It' shrinkable for solaris (ndd -set /dev/tcp tcp_time_wait_interval VALUE_IN_SECONDS), but even there the thread cleaning up the tables runs only every 5 seconds. Concerning the one request 1 connection case: I often realized strange behaviour (unclean shutdown) of ab concerning the last request in a connection. I never analysed it though. If you can easily reproduce the "one request over one connection is slow" problem without high load, you might want to tcpdump to check, if it's really slow on the server side. Just my 0.9 cents ... Rainer Filip Hanik - Dev Lists schrieb: > That's some very good info, it looks like my system never does go over > 30k and cleaning it up seems to be working really well. > btw. do you know where I change the cleanup intervals for linux 2.6 kernel? > > I figured out what the problem was: > Somewhere I have a lock/wait problem > > for example, this runs perfectly: > ./ab -n 1 -c 100 http://localhost:$PORT/run.jsp?run=TEST$i > > If I change -c 100 (100 sockets) to -c 1, each JSP request takes 1 second. > > so what was happening in my test was running 1000 requests over 400 > connections, then invoking 1 request over 1 connection, and repeat. > Every time I did the single connection request, it does a 1sec delay, > this cause the CPU to drop. > > So basically, the NIO connector sucks majorly if you are a single user > :), I'll trace this one down. > Filip > > > Rainer Jung wrote: >> Hi Filip, >> >> the fluctuation reminds me of something: depending on the client >> behaviour connections will end up in TIME_WAIT state. Usually you run >> into trouble (throughput stalls) once you have around 30K of them. They >> will be cleaned up every now and then by the kernel (talking about the >> unix/Linux style mechanisms) and then throughput (and CPU usage) start >> again. >> >> With modern systems handling 10-20k requests per second one can run into >> trouble much faster, than the usual cleanup intervals. >> >> Check with "netstat -an" if you can see a lot of TIME_WAIT connections >> (thousands). If not it's something different :( >> >> Regards, >> >> Rainer >> >> Filip Hanik - Dev Lists schrieb: >> >>> Remy Maucherat wrote: >>> >>>> [EMAIL PROTECTED] wrote: >>>> >>>>> Author: fhanik >>>>> Date: Wed Oct 25 15:11:10 2006 >>>>> New Revision: 467787 >>>>> >>>>> URL: http://svn.apache.org/viewvc?view=rev&rev=467787 >>>>> Log: >>>>> Documented socket properties >>>>> Added in the ability to cache bytebuffers based on number of channels >>>>> or number of bytes >>>>> Added in nonGC poller events to lower CPU usage during high traffic >>>>> >>>> I'm starting to get emails again, so sorry for not replying. >>>> >>>> I am testing with the default VM settings, which basically means that >>>> excessive GC will have a very visible impact. I am testing to >>>> optimize, not to see which connector would be faster in the real world >>>> (probably neither unless testing scalability), so I think it's >>>> reasonable. >>>> >>>> This fixes the paranormal behavior I was seeing on Windows, so the NIO >>>> connector works properly now. Great ! However, I still have NIO which >>>> is slower than java.io which is slower than APR. It's ok if some >>>> solutions are better than others on certain platforms of course. >>>> >>>> >>> thanks for the feedback, I'm testing with larger files now, 100k+ and >>> also see APR->JIO->NIO >>> NIO has a very funny CPU telemetry graph, it fluctuates way to much, so >>> I have to find where in the code it would do this, so there is still >>> some work to do. >>> I'd like to see a nearly flat CPU usage when running my test, but >>> instead the CPU goes from 20-80% up and down, up and down. >>> >>> during my test >>> (for i in $(seq 1 100); do echo -n "$i."; ./ab -n 1000 -c 400 >>> http://localhost:$PORT/104k.jpg 2>1 |grep "Requests per"; done) >>> >>> my memory usage goes up to 40MB, then after a FullGC it goes down to >>> 10MB again, so I wanna figure out where that comes from as well. My >>> guess is that all that data is actually in the java.net.Socket classes, >>> as I am seeing the same results with the JIO connector, but not with >>> APR(cause APR allocates mem using pools) >>> Btw, had to put in the byte[] buffer back into the >>> InternalNioOutputBuffer.java, ByteBuffers are way to slow. >>> >>> With APR, I think the connections might be lingering to long as >>> eventually, during my test, it stop accepting connections. Usually >>> around the 89th iteration of the test. >>> I'm gonna keep working on this for a bit, as I think I am getting to a >>> point with the NIO connector where it is a viable alternative. >>> >>> Filip >>> >>> --------------------------------------------------------------------- >>> To unsubscribe, e-mail: [EMAIL PROTECTED] >>> For additional commands, e-mail: [EMAIL PROTECTED] >>> >> >> --------------------------------------------------------------------- >> To unsubscribe, e-mail: [EMAIL PROTECTED] >> For additional commands, e-mail: [EMAIL PROTECTED] >> >> >> >> > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [EMAIL PROTECTED] > For additional commands, e-mail: [EMAIL PROTECTED] --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
