Re: [PATCH 00/10] Implement batching skb API

Hi Evgeniy,

Evgeniy Polyakov [off-list ref] wrote on 07/20/2007 06:24:23 PM:

Hi Krishna.

On Fri, Jul 20, 2007 at 12:01:49PM +0530, Krishna Kumar

(krkumar2@in.ibm.com) wrote:

quoted

After fine-tuning qdisc and other changes, I modified IPoIB to use this

API,

quoted

and now get good gains. Summary for TCP & No Delay: 1 process improves

for

quoted

all cases from 1.4% to 49.5%; 4 process has almost identical

improvements

quoted

from -1.7% to 59.1%; 16 process case also improves in the range of

-1.2% to

quoted

33.4%; while 64 process doesn't have much improvement (-3.3% to 12.4%).

UDP

quoted

was tested with 1 process netperf with small increase in BW but big
improvement in Service Demand. Netperf latency tests show small drop in
transaction rate (results in separate attachment).

What about round-robin tcp time and latency test? In theory such batching
mode should not change that timings, but practice can show new aspects.

The TCP RR results show a slight impact, however the service demand shows
good improvement. The results are (I did TCP RR - 1 process, 1,8,32,128,512
buffer sizes; and UDP RR - 1 process, 1 byte buffer size) :

        Results for TCR RR (1 process) ORG code:
Size        R-R                   CPU%            S.Demand
------------------------------------------------------------
1         521346.02               5.48            1346.145
8         129463.14               6.74            418.370
32        128899.73               7.51            467.106
128       127230.15               5.42            340.876
512       119605.68               6.48            435.650


        Results for TCR RR (1 process) NEW code (and change%):
Size        R-R                   CPU%            S.Demand
--------------------------------------------------------------------
1         516596.62 (-0.91%)      5.74            1423.819 (5.77%)
8         129184.46 (-.22%)       5.43            336.747 (-19.51%)
32        128238.35 (-.51%)       5.43            339.213 (-27.38%)
128       126545.79 (-.54%)       5.36            339.188 (-0.50%)
512       119297.49 (-.26%)       5.16            346.185 (-20.54%)


              Results for UDP RR 1 process ORG & NEW code:
Code   Size      R-R                CPU%      S.Demand
----------------------------------------------------------------------
ORG     1        539327.86          5.68      1348.985
NEW     1        540669.33 (0.25%)  6.05      1434.180 (6.32%)

I will review code later this week (likely tomorrow) and if there will
be some issues return back.

Thanks! I had just submitted Rev2 on Sunday, please let me know what you
find.

Regards,

- KK