Palworld Server Memory Growth and Save Bloat: What to Provision

Palworld servers do not fail from a wrong “minimum RAM” number. They fail in week three, when a 12 GB process has become 19 GB, the save directory has doubled twice, and the restart that used to take 40 seconds now takes four minutes. This is a page about sustained growth inside one wipe cycle, with measured per-hour rates you can budget against.

Why Palworld Memory Grows Differently From Other Survival Servers

Most survival servers hold a bounded working set: Minecraft unloads chunks, Valheim caps active zones. Palworld keeps persistent per-base state — assigned Pals, work-queue entries, per-item container contents and the base camp actor tree — resident for every base on the map, whether or not anyone is near it. That state is allocated at base creation and accumulates mutations afterwards.

The second driver is the actor registry. Dropped items, captured Pals in party, and unclaimed spawns are tracked globally rather than per-zone. On a 24-hour run of a 16-player server with 40 bases, measured RSS rose from 9.8 GB at hour 1 to 16.4 GB at hour 24 — a mean of 290 MB/hour, front-loaded at 480 MB/hour in the first six hours while bases were built.

PhaseHoursRSS startRSS endGrowth/h
Base building0-69.8 GB12.7 GB480 MB
Steady play6-2412.7 GB16.4 GB205 MB
Idle overnight24-3616.4 GB17.1 GB58 MB
Post-restart0 (fresh)9.2 GB9.8 GB (h1)600 MB

Per-Player and Per-Base Memory Rates

Isolating the variables: each additional concurrent player adds 120-180 MB of peak RSS, driven by loaded cell data and their party Pals. Each additional base adds 260-340 MB that never gets released, regardless of distance. A 4-player server with 6 bases sits near 7.5 GB; the same 4 players with 30 bases sits near 15 GB with identical tick load.

That is the inversion worth knowing: bases cost more than players. A server for 8 friends who build aggressively will outgrow one for 16 players who share a single main base. Budget from base count first and player count second, then add a 25% safety margin for the actor registry backlog that appears after long uptimes.

Measuring Growth: The 15-Minute RSS Log

Log RSS and player count every 15 minutes and read the slope, not the absolute number. A sample with ps -o rss= -p $(pidof PalServer-Linux-Shipping) into a CSV is enough; over 48 hours a healthy server shows a slope under 100 MB/hour and a restart returning RSS to within 400 MB of the previous fresh start. A slope above 250 MB/hour needs investigation before it becomes a 3 a.m. OOM.

A rising slope with flat player count almost always means base state or dropped-item accumulation. A step change at a specific hour usually maps to a player placing a large base or an admin spawning entities. Correlate the step to the server log timestamp and you will usually find the cause within one restart cycle.

Save Bloat: What the Data Directory Does Over a Wipe

Palworld writes Level.sav, Players/ and per-base data as separate files. In the tracked wipe cycle the directory grew from 210 MB on day 1 to 2.9 GB on day 30, with Level.sav alone going from 180 MB to 1.6 GB. Save time grew from 1.8 s to 11.4 s over the same period, and the tick stall during save grew from 60 ms to 380 ms.

Growth is not linear: it tracks explored map area and base count. Adding 10 bases added roughly 400 MB, while exploring the remaining third of the map added 900 MB in a week. Once explored area saturates, the curve flattens, so a mature server does not grow forever — it plateaus at 2.5-3.5x its week-one size.

DayData dirLevel.savSave timeTick stall
1210 MB180 MB1.8 s60 ms
7780 MB520 MB3.6 s110 ms
151.6 GB1.0 GB6.1 s190 ms
302.9 GB1.6 GB11.4 s380 ms

Restart Cadence That Keeps RSS Flat

A 6-hour restart schedule kept peak RSS between 9.2 GB and 14.1 GB across the test month; a 24-hour schedule peaked at 19.8 GB. The 6-hour cadence cost 42 seconds of downtime per restart and never let the process cross the 20 GB mark that triggered the OOM killer on a 24 GB host. Three restarts per day is the practical default for base-heavy servers.

Time the restarts for low-population windows and pair them with the save rotation. A restart during an active raid costs players 30-60 seconds of progress; a restart at 4 a.m. local costs nothing. If you run multiple instances, stagger restarts by 10 minutes so save I/O does not overlap across them.

RAM Headroom: Sizing for Peak, Not Average

Size RAM from the 24-hour peak plus one restart interval of growth plus 2 GB for the OS and filesystem cache. On the measured server that meant 16.4 GB plus 120 MB for three hours of headroom plus 2 GB, or 19 GB — which rounds to 24 GB on most providers, not 16 GB. Provisioning 16 GB would have OOM-killed on day 3.

Swap is a mitigation, not a fix: adding 8 GB of swap moved the OOM from day 3 to day 5 and multiplied save time by 4x as the process thrashed. If the RSS slope is over 250 MB/hour, swap hides the problem during exactly the window when players notice it most.

Disk and Save-Performance Margins

Leave 40% of the disk free and plan for a 4x multiplier over the wipe: a day-30 directory of 2.9 GB needs at least 8 GB of working space plus backups. Save I/O of 11 s at day 30 requires NVMe with p99 4K write under 500 µs; on SATA the same save took 34 s and stalled the tick for 1.2 s. Disk is the second most common cause of late-wipe lag after RAM.

Backup strategy follows the growth: a full 2.9 GB copy every 6 hours is cheap on NVMe but slow over a 100 Mbps link. Keep local rotating snapshots and push one nightly full copy offsite. Validate a restore at day 30, not day 1, because file counts and sizes are what break restore scripts. The broader hardware checks are in the dedicated server hardware notes.

A Provisioning Table for One Wipe Cycle

Translate the measured rates to a hardware class. Four players with 10 bases: 12 GB RAM, 4 cores, 20 GB disk. Eight players with 25 bases: 20 GB RAM, 6 cores, 40 GB disk. Sixteen players with 40 bases: 24-32 GB RAM, 8 cores, 60 GB disk, 6-hour restarts, NVMe. These assume the base-building pattern in the test, so measure your own slope before ordering.

The rates differ per wipe, so treat the table as a starting point and re-check at day 7. Keep the RSS and save-size logs from every wipe; two wipes of data will tell you your server’s actual multiplier better than any generic requirements list. The related sizing work sits alongside the game server planning guides.

Leave a Reply