The Co-Design of Exascale Storage Architectures (CODES) simulation framework builds upon the ROSS parallel discrete event simulation engine to provide high-performance simulation utilities and models for building scalable distributed systems simulations
Other
40
stars
16
forks
source link
DFDally Synthetic: Random Permutation Traffic Determinism and Behavior Fixes #211
This PR fixes a behavioral issue with the random permutation traffic pattern and makes it deterministic.
It fixes the random permutation (--traffic=2) behavior to be in line with what
is expected in a random permutation traffic pattern. The previous implementation
did not actually do any random permutation as there was no counter to determine
how much data had been transmitted to the destination. Consequently there
was no conditional to check for said counter to switch to a new destination.
Each workload rank picked their own random destination and sent all of their
data to it.
The new implementation adds a command line argument --rperm_threshold= which
allows the user to set how much data is transmitted to a specific rank before
a new rperm destination is chosen by the sending workload rank. This should be,
optimally, a multiple of the payload size.
This PR fixes a behavioral issue with the random permutation traffic pattern and makes it deterministic.
It fixes the random permutation (
--traffic=2
) behavior to be in line with what is expected in a random permutation traffic pattern. The previous implementation did not actually do any random permutation as there was no counter to determine how much data had been transmitted to the destination. Consequently there was no conditional to check for said counter to switch to a new destination. Each workload rank picked their own random destination and sent all of their data to it.The new implementation adds a command line argument
--rperm_threshold=
which allows the user to set how much data is transmitted to a specific rank before a new rperm destination is chosen by the sending workload rank. This should be, optimally, a multiple of the payload size.