You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Summary:
The original version would repeat the local device address N times in the payload sent to store during mesh connection rendezvous.
This yields quadratic payload sent to store (N ranks, each sending N).
During later get, there's an outer for loop to grab the remote pair info, discard most of them and just use one (that matches the current rank).
In total this op is cubic which is not efficient for large scale jobs.
In reality if one device is used, for TCP based connection, only the seq number is different, the device addresses are all the same.
This change aims to reduce the payload size sent to store to linear (approx. almost constant -> 1 addr + N seq numbers).
In total this would make the original cubic op to quadratic (or more like O(1.xN) if that makes sense)
Reviewed By: bmaurer
Differential Revision: D45740631
fbshipit-source-id: aa089bfd81f3d392c0aa23a65f497bbacbdf0384
0 commit comments