Skip to content

Commit 69c5438

Browse files
committed
README: Improve README file
1 parent 331ee5d commit 69c5438

File tree

1 file changed

+23
-24
lines changed

1 file changed

+23
-24
lines changed

README.rst

Lines changed: 23 additions & 24 deletions
Original file line numberDiff line numberDiff line change
@@ -252,44 +252,44 @@ them. For other columns, check ``man sacct``.
252252
stripped out and give invalid data. File an issue and this will
253253
be added.
254254

255-
* ``ReqMem``: The raw slurm value in a format like "5Gn". Instead of
256-
parsing this, you probably want to use one of the other values below.
255+
* **Memory related**
257256

258-
* ``ReqMemNode``, ``ReqMemCPU``: Requested memory per node or CPU,
259-
either taken from ReqMem (if it matches) or computed (you might want
260-
to check our logic if you rely on this). In Slurm, you
261-
can request memory either per-node or per-core, and this calculates
262-
the other one for you.
257+
* ``AllocMem``: The ``mem=`` value from ``AllocTRES`` field. You
258+
probably want to use this.
263259

264-
* ``ReqMemType``: ``c`` if the user requested mem-per-core originally,
265-
``n`` if mem-per-node. Extracted from ``ReqMem``. Modern Slurm has
266-
nothing here, and the column value is null.
260+
* ``TotalMem``: The ``mem=`` value from ``TRESUsageInTot`` field.
261+
You probably want to use this.
267262

268-
* ``ReqMemRaw``: The numeric value of the ``ReqMem``, whether it is
269-
``c`` or ``n``.
263+
* ``ReqMem``: The raw slurm value from the ReqMem column.
270264

271-
* ``ReqGPU``: Number of GPUs requested. Extracted from ``ReqTRES``.
265+
* ``ReqMemNode``, ``ReqMemCPU``: Requested memory per node or CPU,
266+
``ReqMem`` / ``NNodes``.
272267

273-
* GPU information. These use values from the ``TRESUsageInAve``
268+
* ``MemEff``: Computed ``TotalMem / AllocMem``.
269+
270+
* **GPU information.** These use values from the ``TRESUsageInAve``
274271
fields in modern Slurm
275272

276-
* ``GpuMem``: ``gres/gpumem``
273+
* ``ReqGPU``: Number of GPUs requested. Extracted from ``ReqTRES``.
274+
275+
* ``GpuMem``: ``gres/gpumem`` from ``TRESUsageInAve``
277276

278277
* ``GpuUtil``: ``gres/gpuutil`` (fraction 0.0-1.0).
279278

280-
* ``NGpus``: Number of GPUs. Should be the same as ``ReqGPU``, but
281-
who knows.
279+
* ``NGpus``: Number of GPUs from ``gres/gpu`` in ``AllocTRES``.
280+
Should be the same as ``ReqGPU``, but who knows.
282281

283282
* ``GpuUtilTot``, ``GpuMemTot``: like above but using the
284283
``TRESUsageInTot`` sacct field.
285284

286-
* ``MemEff``: This is null in the Slurm table now, since Slurm gives
287-
ReqMem in allocations and memory used in steps. The ``eff`` table
288-
calculates this now.
285+
* ``GpuEff``: ``gres/gpuutil`` (from ``TRESUsageInTot``) / (100 *
286+
``gres/gpu`` (from ``AllocTRES``).
289287

290288
* ``CPUEff``: CPU efficiency (0.0-1.0). All the same caveats as above
291289
apply: test before trusting.
292290

291+
* And more, see the code for now.
292+
293293
Quick reference of the other most important columns from the
294294
accounting database that are hardest to remember:
295295

@@ -302,12 +302,11 @@ accounting database that are hardest to remember:
302302

303303
The ``eff`` table adds the following:
304304

305-
* ``CPUEff``: like CPUEff but for the whole job
305+
* ``CPUEff``: Highest CPUEff for any job step
306306

307-
* ``MemEff``: Memory efficiency for the whole job (max(MaxRSS) /
308-
ReqMem)
307+
* ``MemEff``: Highest MemEff for any job step
309308

310-
* And more, see the code for now.
309+
* ``GpuEff``: Highest GpuEff for any job step
311310

312311

313312

0 commit comments

Comments
 (0)