HAIDIS Training Phase
MODEL WEIGHTS
GLUEX DATA 1
10.0.1.10 : 19522
UDP · 10 Gb/s
9000-byte frames
GLUEX DATA 2
10.0.1.11 : 19522
UDP · 10 Gb/s
9000-byte frames
LOAD BALANCER
EJFAT · FPGA
PKT IN
0
PKT OUT
0
DISTRIBUTION
W1
W2
W3
W4
W5
ALGO: ROUND-ROBIN + BACKPRESSURE
ACTIVE
WORKER 1
ERSAP + SAGIPS
10.0.2.10 : 19530
0%
WORKER 2
ERSAP + SAGIPS
10.0.2.11 : 19530
0%
WORKER 3
ERSAP + SAGIPS
10.0.2.12 : 19530
0%
WORKER 4
ERSAP + SAGIPS
10.0.2.13 : 19530
0%
WORKER 5
ERSAP + SAGIPS
10.0.2.14 : 19530
0%
AmSC MLFlow
Model Registry
VERSION
v0
SAVES
0
GlueX Data 1
GlueX Data 2
Model weights (ring)
Ready signal
MPI Rank communications
Model snapshot → registry
t=0.0s