DGX Nightly Benchmark run 20210514

[quasiben]

Benchmark history

Raw Data

<Client: 'tcp://127.0.0.1:45971' processes=10 threads=10, memory=503.79 GiB>
Distributed Version: 2021.04.1+20.gd6f5609a
simple
4.233e-01 +/- 4.273e-02
shuffle
1.996e+01 +/- 9.503e-01
rand_access
5.332e-03 +/- 2.863e-03
anom_mean
1.017e+02 +/- 1.619e+00

Raw Values

simple
[0.49311852 0.38355469 0.43861688 0.45820154 0.41282599 0.43908502
0.35467942 0.44710435 0.35888883 0.44715158]
shuffle
[20.92185233 20.03086835 19.49205927 19.28074589 22.35019035 19.66124184
19.01816171 19.1191214 19.80327522 19.96137403]
rand_access
[0.00775552 0.00311327 0.00431347 0.00349808 0.01038241 0.00346804
0.00378871 0.00320244 0.0105021 0.00329375]
anom_mean
[100.56302905 100.96863079 100.33541369 101.58334684 100.72392011
104.51875496 105.0326643 101.07314181 100.29275966 102.07218337]

Dask Profiles

• Shuffle
  Profile
• Random Access
  Profile
• Simple
  Profile
• Anom
  Mean

Scheduler Execution Graph

[jakirkham]

Run with gc.disable() only as was originally in PR ( #142 ). A follow up run that adds gc.set_threshold(0) is in issue ( #144 )