To repeat the experiments get the code from:
http://reproducingnetworkresearch.wordpress.com/2012/06/09/hedera-2/
or
http://reproducingnetworkresearch.wordpress.com/2012/06/06/hedera/
The paper is straightforward and fairly easy to read.
Things learnt
- VLB does per-flow instead of per-packet spreading of flows to preserve packet ordering. TCP treats unordered packets as a signal of congestion.
- The whole idea of the paper is that Elephant flows must be scheduled/placed carefully to achieve good aggregate bandwidth.
- Global first fit works quite well and there might be no need to run the complicated simulated annealing. The only disadvantage might be the running time of the algorithm.
- NetFPGA Openflow switches have 2 hardware tables: 32 entry TCAM that does matching on variable length prefixes and 32K entry SRAM that matches flow entries with fully qualified 10 tuples
http://reproducingnetworkresearch.wordpress.com/2012/06/09/hedera-2/
or
http://reproducingnetworkresearch.wordpress.com/2012/06/06/hedera/
The paper is straightforward and fairly easy to read.
Things learnt
- VLB does per-flow instead of per-packet spreading of flows to preserve packet ordering. TCP treats unordered packets as a signal of congestion.
- The whole idea of the paper is that Elephant flows must be scheduled/placed carefully to achieve good aggregate bandwidth.
- Global first fit works quite well and there might be no need to run the complicated simulated annealing. The only disadvantage might be the running time of the algorithm.
- NetFPGA Openflow switches have 2 hardware tables: 32 entry TCAM that does matching on variable length prefixes and 32K entry SRAM that matches flow entries with fully qualified 10 tuples
No comments:
Post a Comment