search

bobzhuyb / ns3-rdma

NS3 simulator for RDMA over Converged Ethernet v2 (RoCEv2), including the implementation of DCQCN, TIMELY, PFC, ECN and shared buffer switch
GNU General Public License v2.0
259 stars 118 forks source link

Some question about simulation(about fluid model) #9

Open btlcmr0702 opened 7 years ago

btlcmr0702 commented 7 years ago

hi, yi bo I try to use ns3 to verfiy the fluid model you come up in 'Congestion Control for Large-Scale RDMA Deployments'. Then i get some strange performance. For example, i change parameter BYTECOUNTER to 10MB which comes from your paper, but the rate of host can't converge and queue length at bottleneck varies a lot. Then i found some parameters that i don't understand: CLAMP_TARGET_RATE_ _CLAMP_TARGET_RATE_AFTERTIMER If i set them both to 0, the rate of host can converge but it and queue length at bottleneck still oscillate a lot. I stick the figure of performance(2 flow) as below.

i cyni_haz 1_bg 242ab

Thank you！

bobzhuyb commented 7 years ago

If you are referring to the straight, stable lines in the paper, they are the results of fluid model analysis.

The NS-3 simulation is packet-level. The random marking on switches can't keep the queue as stable as fluid model analysis. As long as your throughput is ~99% link rate and the queue length is mostly ~100KB or less, it is good enough in practice.

btlcmr0702 commented 7 years ago

Ok, thank for your reply. But i still wonder what does these parameter used to? They are in file Qbb-net-device.cc/h _CLAMP_TARGET_RATE CLAMP_TARGET_RATE_AFTERTIMER Will they affect performance a lot?