Skip to content

Latest commit

 

History

History
80 lines (58 loc) · 3.14 KB

decission-support.md

File metadata and controls

80 lines (58 loc) · 3.14 KB
Raw

Decission Support

JULEA can trace backend accesses granularly, allowing further application analysis and object access pattern comparison across applications.

Furthermore, this can be used to provide a first rough estimate of which backend combination is preferable for the application setup.

Access Record

To create an access-record, which is simply a CSV file, execute julea-server with JULEA_TRACE=access and store the output (stderr) in a file. Then run your applications and stop the server when finished.

JULEA_TRACE=access julea-server 2> access-record.csv

Tracing is per default only in debug builds enabled. To build a realease build with trace funtcionality you need to set julea_trace to true

meson configure -Djulea_trace=ture --buildtype=release bld

Replay

An access-record can be replayed with julea-access-replay, which takes only the access-record as input. Replays will execute the same backend operations in the same order as listed in access-record. This allows comparison of backend performance without running the complete application over and over.

Note that the backends must be in the same state, preferably empty, before every replay. Also note that the performance of a replay can vary compared to the original run, depending on the device's load.

JULEA_TRACE=access julea-access-replay access-record.csv 2> replay-record.csv

Setup Testing

The script ./scripts/decission-support.py can help to select a good backend combination for a given workload. It will test different backend configurations and creates a directory evaluation with the results.

Configurations that should be tested must be defined in a config.json file. The expected structure is as follows (an example config can also be found at example/decission-support-config.json):

{
  "object": [{"backend": "posix", "path": "/tmp/<tmp>/posix"}],
  "kv": [{"backend": "sqlite", "path": "/tmp/<tmp>/sqlite"}, {"backend": "sqlite", "path": "/mnt/slow/<tmp>/sqlite"}],
  "db": [{"backend": "sqlite", "path": ":memory:"}],
}

Note that <tmp> inside of paths will be replaced with a temporary directory removed after execution.

The script will replay each access type with all corresponding backends (e.g., all KV accesses are tested on all specified KV configurations). This will produce summary.csv.

The summary can be evaluated using ./scripts/decision-support.R. An example workflow is shown below.

JULEA_TRACE=access julea-server 2> access-record.csv
# run application
# stop julea-server (Ctrl+C)
./scripts/decission-support.py config.json acces-record.csv summary.csv
Rscript ./scripts/decission-support.R summary.csv html > summary.html
# or only ci output
Rscript ./scripts/decission-support.R summary.csv

Important remarks:

  • The script assumes an empty backend state on start.
  • A MySQL/MariaDB instance must be provided if specified as backend.
    • User: julea_user
    • PW: julea_pw
    • Access to the existing (and empty) database julea_db
    • These DB servers must be reset by hand for now. Resets on server startup will be available in the future and the resets will be automated by the script.