Control resources allocated for site

Performance mode toggle

From the site detail page, under Dangerous Actions, you can toggle performance mode for a site on a dedicated server.

How worker allocation works

Changing the performance mode directly affects the number of workers on your bench. Every active bench on a server gets a proportional share of the server's total worker capacity. This runs automatically via auto_scale_workers, which calls allocate_workers on each bench.

Workload score

Each bench gets a workload score: the sum of cpu_time_per_day from the plan of every active, pending, or updating site on that bench. The server's total workload is the sum across all its benches.

The performance mode controls which plan a site is on, and therefore its cpu_time_per_day contribution to the workload score:

• High performance → cpu_time_per_day: 86000
• Not high performance → cpu_time_per_day: 8000

Server RAM budget

Usable RAM is calculated first, then split between the two worker pools:

usable_ram      = max(total_ram - 3000 MB, total_ram × 0.75)
gunicorn pool   = 60% of usable_ram   →  max_gunicorn_workers = pool / 150 MB
background pool = 40% of usable_ram   →  max_bg_workers       = pool / 240 MB

150 MB = avg per gunicorn worker (GUNICORN_MEMORY). 240 MB = avg per set of 3 background workers (BACKGROUND_JOB_MEMORY).

Per-bench allocation formula

Each bench gets a share proportional to its workload, clamped to the release group's configured bounds:

gunicorn_workers   = clamp(min_gn..max_gn,  round(bench_workload / server_workload × max_gunicorn_workers))
background_workers = clamp(min_bg..max_bg,  round(bench_workload / server_workload × max_bg_workers))

These are overridden if the release group has explicit min/max_gunicorn_workers or min/max_background_workers set. If total server workload is zero, every bench falls back to the minimums.

Threaded workers

If gunicorn_threads_per_worker is set on the bench, the gunicorn count is reduced after the proportional step to keep total concurrency the same while using fewer processes:

gunicorn_workers = clamp(1..max_gn, ceil(gunicorn_workers / gunicorn_threads_per_worker))

Memory limits

When set_bench_memory_limits is on, cgroup limits are set per bench:

memory_high = 512 MB + (gunicorn_workers × 150 MB) + (background_workers × 240 MB)
memory_max  = memory_high + 150 MB + 240 MB
memory_swap = memory_max × 2

If skip_memory_limits is set on the bench, it gets the server's maximum possible limits instead. When memory limits are off, all three are set to 0.