Before you start
Use your Linux VPS and its recovery console when SSH is unavailable. These read-only commands do not change the system; some need administrative access or installed tools. Redact sensitive information before sharing output.
Record context before restarting
Note the time zone, affected application and most recent change. Rebooting can remove evidence without resolving the cause. Compare the incident interval with application logs and client-area measurements.
Timestamp and observed load
date -Is
uptimeLoad average is not a CPU percentage: waiting tasks, including tasks waiting for I/O, can raise it. Keep the workload context and do not draw a conclusion from this number alone.
Read available memory rather than cache alone
Linux uses otherwise free memory for reclaimable caches. The available column estimates memory usable by new applications; free and available are not interchangeable. Compare observations taken during the slowdown rather than only immediately after boot.
Memory and swap overview
free -hInspect available alongside RAM and swap totals. Used swap does not prove intensive swapping now. A containerized application can reach its own memory limit while the VPS still has available memory.
Compare swapping, disk waits and pressure
Persistent swapping and waits can accompany latency; merely occupied swap is not a diagnosis. PSI describes time stalled on resource contention, not how much RAM is occupied.
Short system-activity sample
vmstat 1 5The first report summarizes since boot; subsequent reports cover the intervals. Compare si/so, wa and waiting tasks during the incident. An isolated value measures neither network throughput nor an attack.
Memory pressure when the kernel exposes PSI
cat /proc/pressure/memorysome and full describe stalled time, while avg10, avg60 and avg300 cover different windows. A missing file can mean PSI is unavailable; it does not mean pressure is zero.
Look for actual OOM evidence in logs
An application can exit because of a software failure, container limit or maintenance action. Match the service, time and exact event. Without corresponding evidence, report OOM as unconfirmed rather than claim the kernel killed the game server.
Read kernel events from the last hour
journalctl -k --since '-1 hour' --no-pager | grep -Ei 'out of memory|oom-kill|killed process'Additional permissions may be needed. No match does not rule out OOM: the time window, retention or logs may be incomplete. Match the reported process and PID with the affected workload.
Separate SSH listening, CPU and network failures
When the console responds but SSH times out, check local listening state before blaming RAM. A listener does not validate external routing or authentication. If SSH connects but remains slow, compare workload and system activity during the same interval.
TCP listening on the configured SSH port
ss -lnt 'sport = :22'The default port is 22; adapt it only if the service was configured differently. This inspects sockets without opening a port or changing firewall or SSH settings.
Choose a justified correction and measure it
Provide support with timestamps, relevant logs, workload context and authorized external checks. Repeated memory pressure can justify application limits, tuning or more RAM; a routing failure requires a different investigation.
Do not use cache dropping as a universal remedy, delete data or restart repeatedly. Back up before a targeted change, change one element at a time and compare the same workload afterward. A larger plan does not repair software or configuration faults.
For your situation
| Situation | Next check |
|---|---|
| Large cache, available memory and no contention | Do not diagnose exhaustion from cache alone; observe the actual workload. |
| Swapping and pressure coincide with slowdowns | Identify affected processes and limits before changing settings or capacity. |
| SSH times out but the console responds | Compare the listener, address, port and external path rather than automatically blame RAM. |
Official sources
These documents describe the relevant protocols and software. Your service terms and configuration still need to be checked.