[lkml]   [2006]   [Sep]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: Kernel drops ethernet packets during disk writes
Now this is something constructive!

I can indeed maintain 350Mbit/s to disk (about 55MB/s to disk
actually), and I can receive all the packets, as long as I don't try
to do both things simultaniously.

I'll try to increase the buffers tomorrow, will let you know how it
worked out. I doubt though that the tcp buffers have any impact, we're
talking raw ethernet frames here.



On 9/5/06, Stephen Degler < > wrote:
> Tiemen Schut wrote:
> > Summary: The linux kernel appears to drop raw ethernet packets if
> > another process is writing to disk.
> >
> > Test environment: Used a p4 1.7 GHz with gigabit interface
> > point-to-point connection to another p4 (windows pc). This windows PC
> > generates raw ethernet frames holding a counter and sends 'm on to the
> > linux PC, at a transfer rate of 350 Mbit/s. When not writing to disk,
> > everything goes quite fine. I can check the counters at the linux
> > side, and will notice a minimal packet loss (< 0.001 % or so).
> >
> > However, for my application I want to write each and every frame to
> > disk. So, I created a second app, and through a fifo the receiver app
> > and the disk writer app communicate. Now I'm losing like 80% of my
> > packets :o Reducing the throughput on the network doesn't really help
> > (though it does help a little).
> >
> > Note: I tried everything in the same app first, but my guess was that
> > the write operation delayed the app, so I decided to put everything in
> > two apps to give the scheduler some work ;) Didn't work though.
> >
> > My guess would be that the kernel drops the packets because the write
> > operation takes to long (how long can it take, it's just a stupid 1512
> > bytes frame). Anyway, I tried to enlarge
> >, but that didn't do the trick.
> >
> > This problem occures on both 2.6.13 and 2.4.idontremember.
> >
> > It kinda sucks, what's the use of receiving traffic if you can't write
> > it to disk?
> >
> > I'm sending the packets using the winpcap library, and I'm receiving
> > the packets using the pcap library.
> >
> > Any help would be _greatly_ appreciated, and if neccessary, please ask
> > for additional information/used software/test results/etc.
> >
> > Tiemen Schut
> Hi your packets are being dropped perhaps because there is no room to
> store them while the disk is busy, not that you don't receive them.
> Can you sustain 35M/second to disk? If you can't do that with dd
> if=/dev/zero of="foo" bs=1024k, then you can not expect to do it with
> the network plus your writer. Writing
> sequentially to a large modern disk you should be able to do this (large
> capacity SATA or SCSI disks, or hardware raid controllers).
> You can improve matters greatly by increasing the kernel network buffers
> available. This will give you more buffer room while the disk subsystem
> is busy. Look at the following tuneables: you should be able to greatly
> reduce packet loss by increasing them from the default values.
> net.core.rmem_max = 2500000
> net.core.wmem_max = 2500000
> net.ipv4.tcp_rmem = 4096 5000000 5000000
> net.ipv4.tcp_wmem = 4096 65536 5000000
> I can assure you that with the proper system configuration, high network
> to disk throughput is indeed possible.
> Hope this helps,
> skd
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2006-09-05 22:25    [W:0.035 / U:1.144 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site