I can currently get a result of ~60 ops/second. However, this is far
below what this class of machine be receiving. ~600 ops/second would be a
more reasonable number, and even that is on the low end.
The CPU is about 70% idle for most of the test. The client receives
"ECONNREFUSED" when it tries to connect to the box as the test is heating
up.
This test particularly has a very large number of processes in the
"TIME_WAIT" state.
The server machine is a 266 Mhz PII linux box with 256Meg of memory.
It has a FDDI network card, so network bandwidth is not a problem. FDDI
also has the nice property of no collisions, so that also isn't an issue.
OS version: Redhat 5.2
Kernel Version: 2.2.1
Apache Version: 1.3.4
What kind of kernel tunings do I need for optimal performance?
I have changed the values in /usr/src/linux/include/net/tcp.h to:
#define TCP_HTABLE_SIZE 2048 (was 512)
#define TCP_LHTABLE_SIZE 128 (was 32)
#define TCP_BHTABLE_SIZE 2048 (was 512)
I have changed /proc/sys/fs/inode-max to 130712 (was 8192)
I have changed /proc/sys/net/ipv4/ip_local_port_range to 1024 32678
(was 1024 4999)
File-nr does not appear to be a problem. During a test, about ~400 of the
possible 4096 file descriptors are used.
On digital Unix, some of the tunnings used are:
inet: ipqs=7, tcbhashnum=7, tcbhashsize=4096, tcbquicklisten=1,
tcp_sendspace=60000,tcp_recvspace=60000
net: netisrfair=1, netisrwakeupthreshold=256, netisrthreads=3
socket: somaxconn=32767, sominconn=32767, umc=1
proc: max-proc-per-user=1024, maxusers=1024, max-threads-per-user=4096,
per-proc-data-size=0x40000000000, max-per-proc-data-size=0x40000000000,
max-per-proc-address-space=0x4000000000,
per-proc-address-space=0x40000000000
vm: ubc-maxpercent=80, vm-mapentries=50000, vm-vpagemax=131072,
dump-user-pte-pages=1, vm-maxvas=0x40000000000
Does any one know the equivalents for linux?
Thanks,
--Phil
Digital/Compaq: High Performance Servers/Benchmark Performance Engineering
Phillip.Ezolt@compaq.com ezolt@perf.zko.dec.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/