Logo

Platform Lab Annual Review 2021 - Shared screen with speaker view
Hanhong Xue, Google
02:56:00
Did any group try to take advantage of the unfairness?
balaji prabhakar
02:56:43
Thanks Hanhong. I’ll bring up the question at the end of the presentation.
Charif-SGX
03:10:06
How the market data feed will be distribute from match-engine to the trader system in this cloud Ex ? . Multicast /unicast push -pull method ?
Jinkun Geng (Stanford)
03:11:07
unicast
Yuliang Li, Google
03:11:18
If unfairness happens in real world, will there be any solution or compensation, or "it just happened"?
Jinkun Geng (Stanford)
03:11:32
using zmq pub-sub model
Charif-SGX
03:11:36
thanks
Yuliang Li, Google
03:14:04
Thanks!
Aldrin Isaac
03:33:01
Are there NIC vendors that are generally available and support the per-flow pause feature that seems to be needed here? A major NIC supplier I supplier I spoke to says they only support per-flow shaping and not per-flow pause.
Gautam Kumar (Google)
03:33:35
How sensitive is OnRamp's performance is to:1. using RTT instead of OWD2. using rate/cwnd reduction instead of pausing flows
Gautam Kumar (Google)
03:35:49
Thank you!
balaji prabhakar
03:45:23
Gautum: to your question… You could use OR to rapidly detect OWD increase and drastically reduce the rate/window of the CC, but when the transient congestion blows by, you’ll have to increase the rates/windows rapidly again or risk underutilization. OR can’t help during increase.
balaji prabhakar
03:46:21
So, OR is ideally useful for enforcing short-term pause. And the CC takes care of rate/window variations. Hope this makes sense.
Aldrin Isaac
03:55:46
In regards to “spreading across failure domains” what constitutes a good failure domain? What is required in terms of the network capacity across failure domains? What of bandwidth oversubscription?
Nick McKeown
04:00:36
I think you were out by 47ms….
Nick McKeown
04:00:44
Thanks!