This contains a utility, slurm2sql
, which uses the Slurm workload manager's sacct
,
to export statistics from jobs and load them to a well-formed SQLite3
file (the database is also made so that it can be quried with DuckDB).
This file can then be queried for analytics much more easily than the
raw database or your own exports. The main features are:
- Parse
sacct
output (this was made before it had JSON output, which you might want to look at instead - it's hard to use though). - Preprocess all the values in to basic units, including values like GPU usage that currently have to be extracted from other fields.
Even if SQLite isn't what you need, it provides an easy intermediate file on the way to convert to whatever format you want. In particular, it defines the database so that it can be used with DuckDB, which is a more efficient tool for analytics.
There are also some command line frontends, slurm2sql-sacct
and
slurm2sql-seff
that use this parsing to print out data in better
forms than built-in Slurm commands. This is especially useful for
sacct
. You can design your own tools like this.
Normal pip
installation, name slurm2sql
for the command line
programs. This installs the library and command line programs.
pip install slurm2sql
There is only a single file with no depencecies for the core
slurm2sql
library (which could also be manually downloaded - HPC,
right?), though the command line programs require tabulate
. It's
made to support very old Python.
Sample usage:
slurm2sql.py OUTPUT_DB -- [SACCT_FILTER_OPTIONS]
For example, to get all data from July and August (-S
) for all
users (-a
):
slurm2sql.py sincejuly.sqlite3 -- -S 2019-07-1 -a
To get the data from the last N days. This will, day by day, get
each of these history and cumulatively update the database. This
updates a database by default, so that it can be used every day in
order to efficiently keep a running database. The -u
option means
"don't delete existing database" (jobs with the same JobID get
updated, not duplicated):
slurm2sql.py --history-days=N -u sincejuly.sqlite3 -- -a
The --history-start=YYYY-MM-DD
option can do a similar thing
starting from a certain day, and --history=DD-HH:MM:SS
starts
collecting from a given interval of time ago (the time format is as in
Slurm).
To resume from where you left off, first run with one of the history
options. Then, you can do --history-resume
(no -u
needed) and
it will continue fetching day-by-day from the time you last fetched.
You can also run this every day, to first load old historykeep a database updated:
slurm2sql.py --history-days=N -u sincejuly.sqlite3 -- -a slurm2sql.py --history-resume sincejuly.sqlite3 -- -a
This probably isn't the most useful part. Look at command line options.
$ slurm2sql-sacct SACCT_FILTER
This is more useful: it prints seff
like output in a tabular
format. MemReqGiB is per-node, to compare withMaxRSSGiB.
$ slurm2sql-sacct SACCT_FILTER
$ slurm2sql-seff -S now-3day
JobID User hours NCPUS CPUeff MemReqGiB MaxRSSGiB MemEff NGpus GPUeff read_MiBps write_MiBps
------- ------- ----- ----- ------ --------- --------- ------ ----- ------ ---------- -----------
1860854 darstr1 0.28 1 87% 50 9.76 20% 213.88 14.51
1877467 darstr1 0 0 0% 0 0%
1884493 darstr1 0 1 0% 0.49 0 0%
1884494 darstr1 0 1 0% 0.49 0 0%
It can also be used from Python as what is essentially a glorified parser.
db = sqlite3.connect(':memory:')
slurm2sql.slurm2sql(db, ['-S', '2019-08-26'])
# For example, you can then convert to a dataframe:
import pandas as pd
df = pd.read_sql('SELECT * FROM slurm', db)
DuckDB is a lot like SQLite, but column-oriented and optimized for fast processing of data. The main downsides are slow inserts and columns must have consistent data types, but that's the tradeoff we need. Slurm2sql's SQLite database is created with type definitions, so that you can easily open it with DuckDB even without conversion:
$ duckdb dump.sqlite3
Or for even more speed, make a temporary in-memory copy (or this could also be made into a file):
-- command line: $ duckdb database.db
ATTACH ':memory:' AS tmp;
CREATE TABLE tmp.slurm AS (SELECT * FROM slurm);
USE tmp; -- optional but makes tmp the default
Converting to DuckDB:
$ duckdb new.duckdb "CREATE TABLE slurm AS (SELECT * FROM sqlite_scan('original.sqlite3', 'slurm'))"
Using via DuckDB from Python (with the raw sqlite database):
conn = duckdb.connect("database.sqlite3")
conn.execute("select avg(cputime) from slurm").df()
Tables and views:
- Table
slurm
: the main table with all of the data. There is one row for each item returned bysacct
. - View
allocations
: has only the jobs (not job steps) (where JobStep is null
). - View
eff
: Does a lot of processing ofslurm
to produce someCPUEff
,MemEff
, andGPUeff
values (0.0-1.0 usage fractions), in addition to a bit more.
In general, there is one column for each item returned by sacct
,
but some of them are converted into a more useful form. Some columns
are added by re-processing other columns. See COLUMNS
in
slurm2sql.py
for details. Extra columns can easily be added.
Developer note: There are two types of converter functions to make the columns: easy ones, which map one slurm column directly to a database column via a function, and line functions, which take the whole row and can do arbitrary remixing of the data (to compute things like CpuEff.
All column values are converted to standard units: bytes (not MB, KB, etc), seconds, fraction 0.0-1.0 for things like percentages, and unixtime.
Columns which are the same in raw sacct
output aren't documented
specifically here (but note the default units above).
Below are some notable columns which do not exist in sacct (for the
rest, check out the sacct manual page). It's good
to verify that any of our custom columns make sense before trusting
them. For other columns, check man sacct
.
Time
: approximation of last active time of a job. The first of these that exists:End
,Start
,Submitted
. This is intended to be used when you need to classify a job by when it ran, but you don't care to be that specific. (Only the Time column is indexed by default, not the other times)Submit
,Start
,End
: like the sacct equivalents, but unixtime. Assume that the sacct timestamps are in localtime of the machine doing the conversion. (slurm2sql.unixtime
converts slurm-format timestamp to unixtime)QueueTime
is Start-Submit in seconds. Start/End do not include timezones, so expect inaccuracies around summer time changes.Job IDs. Slurm Job ID is by default of format
JobID.JobStep
orArrayJobID_ArrayTaskID.JobStep
. Furthermore, each array job has a "Raw JobID" (different for each job, and is an actual JobID) in addition to the "ArrayJobID" which is the same for all jobs in an array. We split all of these different IDs into the following fields:JobID
: The full raw value that Slurm gives. The same for each job in an array.Only the integer Job ID, without the trailing array tasks or job IDs. For array jobs, this is the "Raw JobID" as described above, use
ArrayJobID
to filter jobs that are the same. IntegerJobIDnostep
: The part of JobID without anything after the.
(no steps)JobIDonly
: The integer part of the JobID.JobIDRawonly
: The integer part of the Raw JobID (so this is different for each job in an aray).ArrayTaskID
: As used above. Integer on null.JobStep
: Job step - only. If you SQL filter forStepID is null
you get only the main allocations. String.Note: HetJob offsets are not currently handled and silently stripped out and give invalid data. File an issue and this will be added.
ReqMem
: The raw slurm value in a format like "5Gn". Instead of parsing this, you probably want to use one of the other values below.ReqMemNode
,ReqMemCPU
: Requested memory per node or CPU, either taken from ReqMem (if it matches) or computed (you might want to check our logic if you rely on this). In Slurm, you can request memory either per-node or per-core, and this calculates the other one for you.ReqMemType
:c
if the user requested mem-per-core originally,n
if mem-per-node. Extracted fromReqMem
. Modern Slurm has nothing here, and the column value is null.ReqMemRaw
: The numeric value of theReqMem
, whether it isc
orn
.ReqGPU
: Number of GPUs requested. Extracted fromReqTRES
.GPU information. These use values from the
TRESUsageInAve
fields in modern SlurmGpuMem
:gres/gpumem
GpuUtil
:gres/gpuutil
(fraction 0.0-1.0).NGpus
: Number of GPUs. Should be the same asReqGPU
, but who knows.GpuUtilTot
,GpuMemTot
: like above but using theTRESUsageInTot
sacct field.
MemEff
: This is null in the Slurm table now, since Slurm gives ReqMem in allocations and memory used in steps. Theeff
table calculates this now.CPUEff
: CPU efficiency (0.0-1.0). All the same caveats as above apply: test before trusting.
Quick reference of the other most important columns from the accounting database that are hardest to remember:
Elapsed
: Wall clock timeCPUTime
: Reserved CPU time (Elapsed * number of CPUs). CPUEff ≈ TotalCPU/CPUTime = TotalCPU/(NCPUs x Elapsed)TotalCPU
: SystemCPU + TotalCPU, seconds of productive work.
The eff
table adds the following:
CPUEff
: like CPUEff but for the whole jobMemEff
: Memory efficiency for the whole job (max(MaxRSS) / ReqMem)- And more, see the code for now.
Next
- This is the biggest column clean-up in a while.
- Add slurm2sql-{seff,sacct} commands.
- JobID columns adjusted:
JobID
is the raw thing that slurm gives,*only
integer IDs without any trailing things,JobIDrawonly
is the RawJobID without any trailing things. - ReqMem has been updated: it no longer parses
n
andc
suffixes for mem-per-node/cpu, and that respective column has been removed. - MemEff has been removed from the
slurm
table, since it is always empty. Theeff
view has been added instead.
0.9.1
- Slurm >= 20.11 deprecates the
AllocGRES
andReqGRES
columns (usingAlloc/ReqTRES
instead).- From this slurm2sql version, a ReqTRES column will be requested and databases will need to be re-created (or manually added to the databases).
- If run on Slurm > 20.11, it will not request ReqGRES and only use ReqTRES.
This could be considered beta right now, but it works and is in use by people. If this is important for you, comment about your use case in the Github issue tracker. Watch the repo if you want to give comments on future data schema updates and backwards compatibility when Slurm changes.
There are many different variations of Slurm, if it doesn't work for you, send an issue or pull request to help us make it more general - development is only done in response to feedback.
Development principles:
- All values are the most basic (metric) units: bytes, seconds, seconds-since-epoch, etc.
- Try to use existing Slurm column names as much as possible (even if they are hard to remember).
- Try to support as many Slurm versions as possible, but if something becomes hard to support, don't worry too much about breaking compatibly. SchedMD support slurm for 18 months after release. Try to support at least those versions. (Until someone asks for it, don't assume we can import data from very old Slurm versions)
- Don't try to maintain database compatibility. It's expected that for all schema changes, you have to delete and re-import. But try to avoid this if not needed.
Release process:
python setup.py sdist bdist_wheel twine upload [--repository-url https://test.pypi.org/legacy/] dist/*0.9.0*
Originally developed at Aalto University, Finland.