search

darshan-hpc / darshan

Darshan I/O characterization tool
Other
56 stars 27 forks source link

strange behavior with DXT MPI-IO heatmap #914

Closed roblatham00 closed 1 year ago

roblatham00 commented 1 year ago

I am digging into an application that reads five HDF5 files. When I turned on DXT, the python report showed an unusual graph for the DXT heatmap. Here's a screen shot of that report:

image

heatmap behaves as I would expect: a bunch of HDF5 reads, followed by a long period of computation and posix HEATMAP vs DXT looks nominal. just eyeballing, it looks like DXT_MPIIO plotted everything in the HEATMAP case but left out the non-io parts. See that big white patch in the heat map case? That's where the app is doing computation. There's no computation in the dxt case. The plot still knows there's 200-some seconds of runtime, but in the dxt case the io calls are scaled weird...to fill up the whole plot.

I've attached a zip file with the darshan log and the generated html.

dxt-mpiio-heatmap-from-hepps.zip

tylerjereddy commented 1 year ago

Reproduced on latest main, I'll try to look into it