GPU allocation

This example creates one candidate row for every compatible pool and workload pair. The model allocates integer GPU units while respecting capacity and demand on both sides.

Input tables

Download gpu_pools.csv and workloads.csv, or create them from the snippets below.

pool_id,capacity,mem_gb,cost
1,8,24,255
2,6,16,180
3,4,32,218

workload_id,demand,min_mem
101,4,8
102,1,8
103,3,32

Build candidate pairs

CREATE CANDIDATES pool_assignments
DECISION KEY (pool_id, workload_id)
AS (
  SELECT
    pool_id,
    workload_id,
    capacity,
    demand,
    mem_gb,
    min_mem,
    cost
  FROM gpu_pools
  CROSS JOIN workloads
  WHERE mem_gb >= min_mem
);

The cross join creates possible assignments. The WHERE clause removes pools without enough memory for a workload.

Allocate GPUs

DECIDE INTO allocation
FROM pool_assignments
DECISION COLUMNS (
  gpus INTEGER BETWEEN 0 AND 6
)
SUBJECT TO
  CONSTRAINT capacity:
    SUM(gpus) <= capacity BY (pool_id),
  CONSTRAINT demand:
    SUM(gpus) = demand BY (workload_id)
MINIMIZE SUM(cost * gpus);

The capacity constraint is repeated for every pool. The demand constraint is repeated for every workload.

Run the model

./deql \
  --register-files gpu_pools.csv,workloads.csv \
  pools_candidates.deql \
  pools_query.deql

The minimum-cost solution places workloads 101 and 102 on pool 2, then serves workload 103 from pool 3.

pool_id  workload_id  capacity  demand  mem_gb  min_mem  cost  gpus
2        101          6         4       16      8        180   4.0
2        102          6         1       16      8        180   1.0
3        103          4         3       32      32       218   3.0

optimal · objective 1554 · highs