[CentOS] TCP offload cards in linux

nate centos at linuxpowered.net
Fri Jun 13 17:52:26 UTC 2008


Anyone have experience with any? I've been having a real hard
time finding info on any cards that actually support this under
linux. (most of the cards work but I don't see drivers that
actually offload the TCP stack)

I have seen some comments where kernel developers don't like the
idea as well.

I'm running CentOS 4.6 64-bit on a dual proc quad core system with
8GB memory.

I have a situation where I'm trying to make the most of my
hardware here, a web application that serves up hundreds of tiny
requests per second(each taking typically sub 100ms to complete).

With my load balancer in it's "normal" mode I can pump about 1000
requests a second through the system, before exhausting the TCP
stack on the server(all ~65k sockets in use). CPU usage tops out
at about 75%. The "normal" mode terminates idle connections
forcefully after 25 seconds.

If I tell my load balancer to terminate idle connections *immediately*
instead of waiting 25 seconds, I get similar CPU usage but transactions
per second drop by about 30-33%.

During my tests I've also gotten 4 kernel panics, the last two
I have captured and they are virtually identical, with the
exception of one time it panic's on the java process, another time
it panics on the swapper process. Both times the load balancer was
configured to terminate connections faster than 25 seconds. I've
done several searches and can't find anything remotely related to
the below panic.

Kernel BUG at tcp_output:943
invalid operand: 0000 [1] SMP
CPU 2
Modules linked in: dell_rbu md5 ipv6 autofs4 i2c_dev i2c_core nfs lockd
nfs_acl sunrpc dm_mirror dm_mod joydev button battery ac uhci_hcd ehci_hcd
hw_random shpchp bnx2 ext3 jbd ata_piix libata megaraid_sas sd_mod scsi_mod
Pid: 10645, comm: java Not tainted 2.6.9-67.ELsmp
RIP: 0010:[<ffffffff802e1a12>] <ffffffff802e1a12>{tcp_retransmit_skb+639}
RSP: 0000:0000010006963e98  EFLAGS: 00010202
RAX: 00000101687cd9c0 RBX: 00000102266d2a80 RCX: 00000101687cd9c0
RDX: 00000101dec72e80 RSI: 0000000000000350 RDI: 0000000000000010
RBP: 000001015c0f7700 R08: 0000000000000008 R09: 0000000000000100
R10: 0000000000000000 R11: 0000000000000008 R12: 000001007b64ccb0
R13: 00000000000005b4 R14: 000001007b64ccb0 R15: 000001007b64c980
FS:  0000000048fe0960(005b) GS:ffffffff804f2e80(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 000000008005003b
CR2: 0000002adc82e028 CR3: 000000000694e000 CR4: 00000000000006e0
Process java (pid: 10645, threadinfo 0000010055cba000, task 000001005633c7f0)
Stack: 0000010006963ec8 0000020000102a93 000001007b64ccb0 0000000000000001
       000001007b64c980 0000000000000008 000001007b64ccb0 000001007b64ccb0
       0000002c273fc4d0 ffffffff802e3f6d
Call Trace:<IRQ> <ffffffff802e3f6d>{tcp_write_timer+1059}
<ffffffff802e3b4a>{tcp_write_timer+0}
       <ffffffff80140945>{run_timer_softirq+356}
<ffffffff8013cff0>{__do_softirq+88}
       <ffffffff8013d099>{do_softirq+49}
<ffffffff80110bf5>{apic_timer_interrupt+133}
        <EOI>

Code: 0f 0b 48 c1 34 80 ff ff ff ff af 03 48 8b 43 10 4c 63 f6 ff
RIP <ffffffff802e1a12>{tcp_retransmit_skb+639} RSP <0000010006963e98>
 <0>Kernel panic - not syncing: Oops

>From what I can tell the Broadcom chips in this server support TOE
but the driver in linux does not implement it. I'd like to find a
card that really has TOE in linux, if for nothing else to compare how
it performs vs what I have now. With the goal of pumping out the
highest number of transactions per second possible with the hardware/power
utilization.

thanks

nate




More information about the CentOS mailing list