On Thu, 26 Oct 2000, Stephen Bailey wrote:
> I think the upshot is that if STPers focus on jumbo frame performance,
> we're going to look like we're completely out of touch with the real
> world.
>
> So (I'm almost afraid to ask) what's the performance for 1500 byte MTU
> (1K STUs)? The good news is that, you'll probably be able to say < 2%
> CPU, and the bad news is that the data rate is going to be, well,
> lower too.
./gen4 starting on toy3 ( reclen=524288 TCPhost=192.168.9.2 ) - Fri Oct 27
14:16:18 2000
# description host sample_KB total_MB sample_KB/s avge_KB/s
cpu_sec user_sec sys_sec sec/MB cpu_pct
1 source toy3 209715.203 209.715 47240.516 47240.516
0.060 0.010 0.050 0.000 1
1 source toy3 209715.203 419.430 47267.035 47253.775
0.090 0.000 0.090 0.000 2
1 source toy3 209715.203 629.146 47297.066 47268.206
0.040 0.000 0.040 0.000 1
2k STU's give around 80MB and 4k is over 100.
The hardware used is a dual pIII/500 with 66MHz PCI sending to a pII/400
with standard frames, the other way around the performance is slightly
lower.
To give a comparison, TCP gets about 62M/s with ~= 100% CPU on my
hardware.
As for latency,
[root@toy3 i686-linux]# ./lat_stp 192.168.9.2
STP latency using 192.168.9.2: 569.2200 microseconds
[root@toy3 i686-linux]# ./lat_tcp 192.168.9.2
TCP latency using 192.168.9.2: 149.6932 microseconds
Which isn't too bad considering STP needs that extra round-trip for
every write() to setup the buffers.
|