We saw something similar to this but it was the "rpciod" task. I don't
know if it applies or not in your situation, but it seemed to help tcpip
performance in general so I thought I would share.
The solution for us was two fold.
1.) Increase some Linux kernel buffers. These notes are from the netapp
website (now.netapp.com) since they are specifically interested in NFS
tuning.
Enlarging the transport socket buffers your client uses for NFS traffic
helps reduce resource contention on the client, reduces performance
variance, and improves maximum data and request throughput. In future
releases of the client, this procedure won't be necessary, as the client
will automatically choose an optimal socket buffer size.
1. Become root on your client
2. cd into /proc/sys/net/core
3. echo 262143 > rmem_max
4. echo 262143 > wmem_max
5. echo 262143 > rmem_default
6. echo 262143 > wmem_default
7. Remount your NFS file systems on the client
This is especially useful for NFS over UDP and when using Gigabit Ethernet.
You should consider adding this to a system startup script that runs before
the system mounts NFS file systems. The size we recommend is the largest
safe socket buffer size we've tested. On clients smaller than 16MB, you
should leave the default socket buffer size setting to conserve memory.
Follow these instructions to ensure that enough memory is reserved for TCP
sockets on large systemsg:
1. Become root on your client
2. cd into /proc/sys/net/ipv4
3. echo "4096 87380 8388608" > /proc/sys/net/ipv4/tcp_rmem
4. echo "4096 87380 8388608" > /proc/sys/net/ipv4/tcp_wmem
5. Remount your NFS file systems on the client
Other TCP Enhancements
The following settings can help reduce the amount of work your clients and
filers do when running NFS over TCP:
1. echo 0 > /proc/sys/net/ipv4/tcp_sack
2. echo 0 > /proc/sys/net/ipv4/tcp_timestamps
These disable optional features of the TCP protocol to save a little
processing and network bandwidth. If you build your own kernels, be sure
that CONFIG_SYNCOOKIES is disabled. SYN cookies slow down TCP connections
by adding extra processing on both ends of the socket. Some Linux
distributors provide kernels with SYN cookies enabled.
Linux 2.2 and 2.4 kernels support large TCP windows (RFC 1323) by default.
No modification is needed to enable large TCP windows.
2.) Increase packet / MTU sizes.
- Since our primary interest was 8K blocks of data going back and forth to
the netapp filer, we went to 9K "jumbo" frames and that helped some too.
-------------------------------------------
Jeremy Warren
Sr. Systems Programmer
KB Toy Stores
100 West St.
Pittsfield, MA 01201
|---------+---------------------------->
| | Michael Morgan |
| | <michael.morgan@o|
| | racle.com> |
| | Sent by: Linux on|
| | 390 Port |
| | <[EMAIL PROTECTED]|
| | IST.EDU> |
| | |
| | |
| | 10/21/02 01:15 AM|
| | Please respond to|
| | Linux on 390 Port|
| | |
|---------+---------------------------->
>-----------------------------------------------------------------------------------------------------------------------------------|
|
|
| To: [EMAIL PROTECTED]
|
| cc:
|
| Subject: [LINUX-390] ksoftirqd_CPU0
|
>-----------------------------------------------------------------------------------------------------------------------------------|
We have noticed that a process called ksoftirqd_CPU0 is using almost all
CPU on our SuSE 2.4.7 kernel SLES7 system. Has anyone else seen this
before? We are having QDIO OSA-E GB problems and this process is
suspect.
..thanks
