bcronje at gmail
Oct 29, 2009, 3:30 AM
Post #2 of 8
(1843 views)
Permalink
|
Hi Fedor,
Out of interest sake, what are the stats for pf_ring 3.x ? As per a previous
post by Luca non-transparent mode has not been implemented in 4.0 yet, so
the kernel is processing the packets as well which will add to CPU load.
Also what are your interrupt coalescence settings, NIC RX ring size and
pf_ring size?
Beyers
On Thu, Oct 29, 2009 at 12:07 PM, Fedor Sakharov
<fedor.sakharov [at] gmail>wrote:
> Hello.
>
> I have been testing libpcap and pf_ring recently. And the results have
> been significantly different from those mentioned in pf_ring papers. I
> have tested CPU load, interface drop and a packet capture drop.
>
> Hardware used:
>
> three SUN x2250 servers with
>
> CPU: Intel(R) Xeon(R) CPU X5460 @ 3.16GHz
> Memory: 8Gb
> Ethernet controller: Intel Corporation 82571EB Gigabit Ethernet
> Controller (rev 06)
>
> Software used:
>
> Kernel: 2.6.26-amd64
> PF_RING: 4.0 (compiled as a kernel module)
> libpcap: 0.8
> libpcap+mmap: 0.9.8
>
> first node runs PF_RING
> second node runs libpcap
> and the third node runs libpcap+mmap.
>
> Tests carried out:
> I have been generating traffic with pktgen tool with different speeds
> and packet sizes. Here are the results of tests with 64 byte packets.
> The difference in performance between libpcap and libpfring is huge.
>
> libpcap:
> 55 Mbits/sec -- 50% CPU load, 0% dropped by the interface, 0% dropped
> by libpcap.
> 77 Mbits/sec -- 60% CPU load, 0% dropped by the interface, <1% dropped
> by libpcap.
> 94 Mbits/sec -- 66% CPU load, 0% dropped by the interface, <1% dropped
> by libpcap.
> 140 Mbits/sec -- 80% CPU load, 0% dropped by the interface, <1%
> dropped by libpcap.
> 406 Mbits/sec -- 110% CPU load, 4% dropped by the interface, 2%
> dropped by libpcap.
> 726 Mbits/sec -- 160% CPU load, 28% dropped by the interface, 10%
> dropped by libpcap.
>
> libpfring:
> 55 Mbits/sec --12% CPU load, 0% dropped by the interface, 0% dropped
> by libpfring.
> 77 Mbits/sec -- 22% CPU load, 0% dropped by the interface, 0% dropped
> by libpfring.
> 94 Mbits/sec -- 35% CPU load, 0% dropped by the interface, 0% dropped
> by libpfring.
> 140 Mbits/sec -- 141% CPU load, 22% dropped by the interface, 0%
> dropped by libpfring.
> 406 Mbits/sec -- 153% CPU load, 71% dropped by the interface, <1%
> dropped by libpfring.
> 726 Mbits/sec -- 170% CPU load, 83% dropped by the interface, 1%
> dropped by libpfring.
>
> As you see the results differ a lot even on 140 Mbits/sec and when
> pf_ring is used interface drops and CPU load are huge. Is there
> anything that can be done to solve this problem?
>
> --
> Best regards,
> Fedor Sakharov
> _______________________________________________
> Ntop-misc mailing list
> Ntop-misc [at] listgateway
> http://listgateway.unipi.it/mailman/listinfo/ntop-misc
>
|
