Thread (28 messages) 28 messages, 6 authors, 2009-09-11

Re: [PATCH 00/14] pktgen update for net-next (2.6.32)

From: Jesper Dangaard Brouer <hidden>
Date: 2009-09-08 12:21:12

On Tue, 8 Sep 2009, robert@herjulf.net wrote:
Stephen Hemminger writes:
quoted
The biggest change is switching to monotonic clock (ktime) and
high resolution timers for the interpacket delay.
Opteron Barcelona 2.3 GHz with net-next 2.6.31-rc5bifrost-x86_64.

Sending different pkt-sizes at maxrate using one CPU-core (of eight)
with ixgbe driver and Intel's 82598 chip.

     w. patch
64    4761904    3846153
128   4545454    3703703
256   4545454    3703703
512   2380952    2380952
1024  1204819    1204819
1500   826446     826446

We increase max sending rate from 3.8 Mpps to 4.8 Mpps for 64 byte pkts
I can also measure a performance improvement.

Generator machine, AMD Phenom 9950 quad core, using a 10GbE Intel 82599 
NIC (6 port 10GbE from Hotlava Systems Inc.).

With patches:
-------------
tx_pkt_sz: 64  TX-pps: 9426724
tx_pkt_sz: 128 TX-pps: 8220396
tx_pkt_sz: 256 TX-pps: 4463415
tx_pkt_sz: 384 TX-pps: 3063025
tx_pkt_sz: 512 TX-pps: 2331636
tx_pkt_sz: 640 TX-pps: 1879640
tx_pkt_sz: 768 TX-pps: 1577957
tx_pkt_sz: 896 TX-pps: 1359122
tx_pkt_sz: 1024 TX-pps: 1193118
tx_pkt_sz: 1152 TX-pps: 1062726
tx_pkt_sz: 1280 TX-pps: 958408
tx_pkt_sz: 1408 TX-pps: 873167
tx_pkt_sz: 1514 TX-pps: 812591

Without patches:
----------------
tx_pkt_sz: 64  TX-pps: 8372361
tx_pkt_sz: 128 TX-pps: 7915717
tx_pkt_sz: 256 TX-pps: 4462400
tx_pkt_sz: 384 TX-pps: 3059712
tx_pkt_sz: 512 TX-pps: 2332164
tx_pkt_sz: 640 TX-pps: 1882343
tx_pkt_sz: 768 TX-pps: 1577543
tx_pkt_sz: 896 TX-pps: 1358061
tx_pkt_sz: 1024 TX-pps: 1192788
tx_pkt_sz: 1152 TX-pps: 1062959
tx_pkt_sz: 1280 TX-pps: 958622
tx_pkt_sz: 1408 TX-pps: 872934
tx_pkt_sz: 1514 TX-pps: 812771

It really makes a difference for small packet sizes.

  64 bytes pkts: 8.4 Mpps to 9.4 Mpps.
128 bytes pkts: 7.9 Mpps to 8.2 Mpps

Notice that the numbers for pkt sizes above 128 bytes are really close to 
wirespeed 10GbE rates, when taking ethernet framing into account.
Thanks a lot Stephen.
Yes, thanks a lot Stephen! :-)

Cheers,
   Jesper Brouer

--
-------------------------------------------------------------------
MSc. Master of Computer Science
Dept. of Computer Science, University of Copenhagen
Author of http://www.adsl-optimizer.dk
-------------------------------------------------------------------
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help