* impressive throughput on 2.5.73
@ 2003-07-03 0:56 Dave Hansen
2003-07-03 2:58 ` Nivedita Singhvi
0 siblings, 1 reply; 3+ messages in thread
From: Dave Hansen @ 2003-07-03 0:56 UTC (permalink / raw)
To: netdev; +Cc: Scott Feldman, Nivedita Singhvi
I run a little script to load up a gigabit ethernet link between two of
my machines: an 8-way PIII web server, and a 4-way PIII client. Both
client and server have e1000s. The script fetches a bunch of fairly big
files via http from the server, using apache. I'm impressed that it can
fill just about the entire gigabit pipe, while using less than 1 of the
server's CPUs.
176 requests/sec - 110.0 MB/second - 0.6 MB/request
Server CPU breakdown.
2% system
7% user
91% idle
I'm not using any interrupt mitigation, so I'm still at ~9k
interrupts/sec. I'll try NAPI next.
server profile:
197212 total 0.1159
175783 poll_idle 2441.4306
1474 alloc_skb 6.8241
1281 skb_release_data 8.6554
1202 skb_clone 3.7563
1038 do_tcp_sendpages 0.3655
998 e1000_clean_tx_irq 2.1886
974 e1000_xmit_frame 0.5148
841 tcp_transmit_skb 0.5729
827 Letext 2.0675
791 __kmalloc 6.3790
739 ip_queue_xmit 0.6415
629 ip_finish_output 1.4976
583 tcp_write_xmit 0.7836
571 tcp_v4_rcv 0.2896
422 schedule 0.2733
417 e1000_intr 3.7232
403 tcp_clean_rtx_queue 0.5140
399 __kfree_skb 2.1685
353 kfree 3.6771
337 kmem_cache_free 4.4342
309 e1000_clean_rx_irq 0.3053
291 find_get_page 6.0625
284 do_softirq 1.4200
246 eth_type_trans 1.4643
233 dev_queue_xmit 0.4380
226 __wake_up 5.1364
216 ip_rcv 0.2093
202 sock_wfree 3.3667
202 memcpy 5.0500
client profile:
33363 total 0.0196
9575 poll_idle 132.9861
5099 __copy_user_intel 32.6859
3307 schedule 2.1418
1753 __wake_up 39.8409
667 tcp_v4_rcv 0.3382
577 pipe_write 0.7970
510 __down_wq 1.7708
488 alloc_skb 2.2593
476 tcp_recvmsg 0.2216
457 tcp_rcv_established 0.2746
457 __kfree_skb 2.4837
405 dnotify_parent 4.5000
382 pipe_read 0.7290
368 system_call 8.3636
350 kill_fasync 7.0000
341 current_kernel_time 5.0147
299 __kmalloc 2.4113
280 ip_rcv 0.2713
252 eth_type_trans 1.5000
245 skb_release_data 1.6554
233 ip_queue_xmit 0.2023
232 kfree 2.4167
228 e1000_clean_rx_irq 0.2253
210 pipe_wait 1.3816
205 e1000_clean_tx_irq 0.4496
200 tcp_transmit_skb 0.1362
199 e1000_xmit_frame 0.1052
--
Dave Hansen
haveblue@us.ibm.com
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: impressive throughput on 2.5.73
2003-07-03 0:56 impressive throughput on 2.5.73 Dave Hansen
@ 2003-07-03 2:58 ` Nivedita Singhvi
2003-07-03 3:20 ` Dave Hansen
0 siblings, 1 reply; 3+ messages in thread
From: Nivedita Singhvi @ 2003-07-03 2:58 UTC (permalink / raw)
To: Dave Hansen; +Cc: netdev, Scott Feldman
Dave Hansen wrote:
> 176 requests/sec - 110.0 MB/second - 0.6 MB/request
>
> Server CPU breakdown.
> 2% system
> 7% user
> 91% idle
>
> client profile:
> 33363 total 0.0196
> 9575 poll_idle 132.9861
> 5099 __copy_user_intel 32.6859
> 3307 schedule 2.1418
> 1753 __wake_up 39.8409
> 667 tcp_v4_rcv 0.3382
Nice numbers :). In addition to NAPI, would be nice to
stick in more adapters. Any idea why __wake_up should be
so expensive on the client side?
thanks,
Nivedita
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: impressive throughput on 2.5.73
2003-07-03 2:58 ` Nivedita Singhvi
@ 2003-07-03 3:20 ` Dave Hansen
0 siblings, 0 replies; 3+ messages in thread
From: Dave Hansen @ 2003-07-03 3:20 UTC (permalink / raw)
To: Nivedita Singhvi; +Cc: netdev, Scott Feldman
On Wed, 2003-07-02 at 19:58, Nivedita Singhvi wrote:
> Nice numbers :). In addition to NAPI, would be nice to
> stick in more adapters. Any idea why __wake_up should be
> so expensive on the client side?
NAPI seems to degrade throughput down to ~80MB/sec, but the interrupt
rate doesn't change. I'm going to put some debugging in and makes sure
that it's kicking in properly.
--
Dave Hansen
haveblue@us.ibm.com
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2003-07-03 3:20 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2003-07-03 0:56 impressive throughput on 2.5.73 Dave Hansen
2003-07-03 2:58 ` Nivedita Singhvi
2003-07-03 3:20 ` Dave Hansen
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).