2 messages, 2 authors, 2004-08-03 · open the first message on its own page
From: Francois Romieu <romieu@fr.zoreil.com>
Date: 2004-08-03 07:51:48
Bart Alewijnse [off-list ref] : [...]
The panic looks a lot like the last one; same kernel (napi still enabled for the 8169). Image attached.
The irq rate are strangely high for a napi version of the r8169 driver. Can you describe your test commands so that I reproduce these here ? -- Ueimor
From: Bart Alewijnse <hidden>
Date: 2004-08-03 12:32:07
Quite simply a 'netio -s' on one computer, and a 'netio -u 192.168.1.whichever' on the other. I was monitoring with 'vmstat 1' for hardware stuff and bmon to see the actual speed and had a ssh open. The things I've been fiddling with recently to see if it'd make a difference: - Renicing of ksoftirqc to anything between -5 and -17, I don't remember exactly. - Disabling tcp window scaling. I read somewhere that that may help throughhput, but it may be a stupid move. - Some barely thought about (I did read the following: http://lists.samba.org/archive/samba/2003-December/077198.html) window memory buffer size changes, with a 'so that that at least won't be a bottleneck' angle: -------------------- echo 400000 > /proc/sys/net/core/wmem_default echo 512000 > /proc/sys/net/core/wmem_max echo 400000 > /proc/sys/net/core/rmem_default echo 512000 > /proc/sys/net/core/rmem_max echo 98304 512000 640000 > /proc/sys/net/ipv4/tcp_mem echo 98304 512000 640000 > /proc/sys/net/ipv4/tcp_wmem echo 98304 512000 640000 > /proc/sys/net/ipv4/tcp_rmem echo 98304 512000 640000 > /proc/sys/net/ipv4/tcp_mem ----------- It seems that windows (in which I have jumbo frames on) doesn' quite need as many interrupts, although I'm not sure that was for the same speed, as I had another kernel panic (attached, this one has a small visible trace). I was running the windows version of netio as a server, and doing a client test with udp. --Bart Alewijnse