A perfSONAR User's Group

Switch buffer size

I gave a lightning talk in 2010 at Joint Techs that reviewed measurements that show empirically how buffer size affects bwctl throughput. I partitioned the buffer RAM on a Catalyst 6500 Gig-E port to different queues. I set the RED parameters to get pure tail drop performance. The bwctl tests were done to the best effort queue. Moving memory in/out of that queue affects TCP throughput.. See the slides at:


So, how much buffer should you expect and is there anything you can do about it? I surveyed the field a while back and attempted to collect buffer info.


If you have info you’d like added to the table, you can send it to me. If you have reason to believe that some of this needs correction, that would be good to know.

Comments are currently closed.