"PG-Strom"をコンテナの中で実行する試みです。
- ホストOS: Red Hat Enterprise Linux 8.8もしくは9.2以降
- Docker engine 26.1.1以降
- NVIDIA Container Toolkit の要件
- NVIDIA Container Toolkit をセットアップ
- Dockerfileを記述
例えば...
- コンテナイメージを作成
# cd pg-strom-docker/docker
# docker image build --compress -t mypg15-rocky8:latest -f Dockerfile .- コンテナを起動
--shm-size オプションを使用して、適切な共有メモリをコンテナーに設定します。
# docker container run --gpus all --shm-size=8gb --memory=8gb -p 5432:5432 -itd --name=cont1 mypg16-rocky8:latest
# docker container exec -it cont1 /bin/bash- initdbコマンドをコンテナで実行
# su - postgres
$ /usr/pgsql-16/bin/initdb -D /var/lib/pgsql/16/data- パラメータの変更
例えば...
$ vi /var/lib/pgsql/16/data/postgresql.conf
...
shared_preload_libraries = '$libdir/pg_strom'
max_worker_processes = 100
shared_buffers = 4GB
work_mem = 1GB- PostgreSQL Serverを起動
$ /usr/pgsql-15/bin/pg_ctl -D /var/lib/pgsql/16/data -l logfile start
$ cat /var/lib/pgsql/logfile
2024-05-07 08:12:30.823 UTC [148] LOG: HeteroDB Extra module loaded [api_version=20240418,cufile=off,nvme_strom=off,githash=3ffc65428c07bb3c9d0e5c75a2973389f91dfcd4]
2024-05-07 08:12:30.823 UTC [148] LOG: PG-Strom version 5.12.el8 built for PostgreSQL 16 (githash: )
2024-05-07 08:12:32.481 UTC [148] LOG: PG-Strom binary built for CUDA 12.4 (CUDA runtime 12.4)
2024-05-07 08:12:32.481 UTC [148] LOG: PG-Strom: GPU0 NVIDIA GeForce GTX 1650 SUPER (20 SMs; 1725MHz, L2 1024kB), RAM 3720MB (128bits, 5.72GHz), PCI-E Bar1 0MB, CC 7.5
2024-05-07 08:12:32.482 UTC [148] LOG: [0000:01:00:0] GPU0 (NVIDIA GeForce GTX 1650 SUPER; GPU-44a7564d-a1a1-bc4e-699a-db8cec30ee7d)
2024-05-07 08:12:32.482 UTC [148] LOG: [0000:02:00:0] nvme0 (CL1-3D256-Q11 NVMe SSSTC 256GB) --> GPU0 [dist=5]
2024-05-07 08:12:32.522 UTC [148] LOG: redirecting log output to logging collector process
2024-05-07 08:12:32.522 UTC [148] HINT: Future log output will appear in directory "log".- 試してみましょう
# su - postgres
Last login: Wed Nov 9 02:52:33 UTC 2022 on pts/0
$ psql
psql (16.2)
Type "help" for help.
postgres=# CREATE DATABASE testdb2;
CREATE DATABASE
postgres=# \c testdb2
You are now connected to database "testdb2" as user "postgres".
testdb2=#
testdb2=# CREATE extension pg_strom;
CREATE EXTENSION
testdb2=# CREATE TABLE t_test AS SELECT id, id % 10 AS ten, id % 20 AS twenty
FROM generate_series(1, 25000000) AS id
ORDER BY id;
testdb2=# CREATE TABLE t_join AS SELECT *
FROM t_test
ORDER BY random()
LIMIT 1000000;
testdb2=# VACUUM FULL t_test;
testdb2=# VACUUM FULL t_join;
testdb2=# SELECT pg_size_pretty(pg_relation_size('t_test'));
pg_size_pretty
----------------
1056 MB
(1 row)
testdb2=# SELECT pg_size_pretty(pg_relation_size('t_join'));
pg_size_pretty
----------------
42 MB
(1 row)
testdb2=# SET pg_strom.enabled=off;
testdb2=# explain analyze SELECT count(*) FROM t_test AS a, t_join AS b WHERE a.id = b.id GROUP BY a.ten;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------
-------------------------------
Finalize GroupAggregate (cost=389641.98..389692.65 rows=200 width=12) (actual time=2039.033..2044.305 rows=10 loops=1)
Group Key: a.ten
-> Gather Merge (cost=389641.98..389688.65 rows=400 width=12) (actual time=2039.027..2044.297 rows=30 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Sort (cost=388641.96..388642.46 rows=200 width=12) (actual time=2037.315..2037.317 rows=10 loops=3)
Sort Key: a.ten
Sort Method: quicksort Memory: 25kB
Worker 0: Sort Method: quicksort Memory: 25kB
Worker 1: Sort Method: quicksort Memory: 25kB
-> Partial HashAggregate (cost=388632.32..388634.32 rows=200 width=12) (actual time=2037.303..2037.306 rows=1
0 loops=3)
Group Key: a.ten
Batches: 1 Memory Usage: 40kB
Worker 0: Batches: 1 Memory Usage: 40kB
Worker 1: Batches: 1 Memory Usage: 40kB
-> Parallel Hash Join (cost=14781.00..336548.98 rows=10416667 width=4) (actual time=72.352..2002.739 ro
ws=333333 loops=3)
Hash Cond: (a.id = b.id)
-> Parallel Seq Scan on t_test a (cost=0.00..239302.67 rows=10416667 width=8) (actual time=0.044.
.425.452 rows=8333333 loops=3)
-> Parallel Hash (cost=9572.67..9572.67 rows=416667 width=4) (actual time=71.343..71.344 rows=333333 loops=
3)
Buckets: 1048576 Batches: 1 Memory Usage: 47392kB
-> Parallel Seq Scan on t_join b (cost=0.00..9572.67 rows=416667 width=4) (actual time=0.010..24.377
rows=333333 loops=3)
Planning Time: 0.077 ms
Execution Time: 2044.331 ms
(23 rows)
testdb2=# SET pg_strom.enabled=on;
testdb2=# explain analyze SELECT count(*) FROM t_test AS a, t_join AS b WHERE a.id = b.id GROUP BY a.ten;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------
-----------------------
HashAggregate (cost=82655.55..82657.55 rows=200 width=12) (actual time=921.113..921.163 rows=10 loops=1)
Group Key: a.ten
Batches: 1 Memory Usage: 40kB
-> Gather (cost=82632.55..82654.55 rows=200 width=12) (actual time=921.099..921.153 rows=10 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Parallel Custom Scan (GpuPreAgg) on t_test a (cost=81632.55..81634.55 rows=200 width=12) (actual time=905.828..9
05.831 rows=3 loops=3)
GPU Projection: pgstrom.nrows(), a.ten
GPU Join Quals [1]: (a.id = b.id) ... [plan: 10416670 -> 10416670, exec: 25000000 -> 1000000]
GPU Outer Hash [1]: a.id
GPU Inner Hash [1]: b.id
-> Parallel Seq Scan on t_join b (cost=0.00..9572.67 rows=416667 width=4) (actual time=0.008..22.373 rows=333
333 loops=3)
Planning Time: 0.116 ms
Execution Time: 921.749 ms
(14 rows)