Linux Swap Usage on Web Servers: Measurement, Tuning, and the RAM Problem Behind It

How Does Linux Swap Work and Why Does It Slow Web Servers?

Swap is disk space the kernel uses as an overflow area for RAM pages that are not currently needed. When a process needs more memory than is available in RAM, the kernel selects pages that have not been accessed recently (cold pages) and writes them to swap, freeing RAM for the process’s new allocation. When those swapped-out pages are needed again, the kernel reads them back from disk. On a spinning HDD, this swap read takes 5-20ms per page. On SSD, it takes 0.1-0.5ms. PHP process pages being swapped to HDD means PHP execution time increases from 100ms to 3 seconds — because every memory access for a swapped-out page triggers a disk read. The server does not crash (unlike OOM), but it slows to a crawl. The key metric is swap I/O rate, not swap usage. A server with 500MB of swap used but zero swap I/O is fine — the 500MB of pages are sitting on disk untouched (pages that were paged out during a quiet period and have not been needed since). A server with 100MB of swap used but high si (swap in) and so (swap out) in vmstat is actively swapping — those pages are being read and written continuously, causing disk I/O and process slowdowns. vm.swappiness controls how aggressively the kernel uses swap. Value 0-100: low values (1-10) make the kernel exhaust all available RAM before using swap; high values (60-100) cause the kernel to swap out idle pages proactively, keeping more RAM free for file system cache. For web servers, vm.swappiness=10 is standard — avoid swapping unless RAM is genuinely exhausted.

Tools and Commands

# Check current swap usage
free -h
swapon --show

# Monitor swap I/O rate (si=swap in, so=swap out, in KB/s)
vmstat 2 10
# si and so should be 0 almost all the time
# si > 0: pages being read from swap back into RAM (process was swapped, now active)
# so > 0: pages being written from RAM to swap (RAM pressure)

# Check vm.swappiness
cat /proc/sys/vm/swappiness

# Set vm.swappiness temporarily
sysctl -w vm.swappiness=10

# Persist across reboots
echo 'vm.swappiness = 10' >> /etc/sysctl.d/99-performance.conf
sysctl -p /etc/sysctl.d/99-performance.conf

# Identify which processes have pages in swap
for pid in $(ls /proc | grep '^[0-9]'); do
  swap=$(awk '/VmSwap/{print $2}' /proc/$pid/status 2>/dev/null)
  comm=$(cat /proc/$pid/comm 2>/dev/null)
  [[ $swap -gt 0 ]] && echo "${swap}kB $comm (PID $pid)"
done 2>/dev/null | sort -rn | head -15

vmstat’s si/so columns are the most important swap metrics for performance — they tell you if swapping is actively happening right now. The /proc loop shows which processes have pages currently on swap, ranked by swap usage.

Key Parameters

Diagnosis and Fix Workflows

Determine if swap I/O is causing current slowness

Run vmstat and check si/so columns. If they are consistently above 0 during a performance complaint, swap I/O is contributing to the slowdown. Correlate with disk iostat await for the swap device.

# Run vmstat with 1-second intervals, watch si and so columns
vmstat 1 30

# Column layout:
# procs: r b
# memory: swpd free buff cache
# swap: si so
# io: bi bo
# system: in cs
# cpu: us sy id wa st

# Look for:
# swpd > 0 AND (si > 0 OR so > 0) = active swapping
# wa (cpu wait) > 20% correlating with si > 0 = swap I/O causing CPU wait

# Cross-reference with iostat to see which device is handling swap I/O
iostat -xz 1 10 | grep -E 'Device|sda|nvme'

Reduce swap usage by identifying and fixing the RAM consumer

Active swapping means RAM is exhausted. Find the processes consuming the most RAM, then reduce their allocation or the number of instances.

# Find top RAM consumers
ps --no-headers -eo rss,pid,comm --sort=-rss | head -10 | awk '{printf "%.0fMB  %s (PID %s)\n", $1/1024, $3, $2}'

# Compare to memory budget:
free -m
# If used >> available: over-provisioned processes

# Most common over-provisioned services on LAMP:
# 1. Apache prefork MaxRequestWorkers too high:
apachectl -V | grep MPM
grep MaxRequestWorkers /etc/apache2/mods-available/mpm_prefork.conf

# 2. PHP-FPM pm.max_children too high:
grep pm.max_children /etc/php/8.1/fpm/pool.d/www.conf

# 3. MySQL innodb_buffer_pool_size too large:
mysql -e 'SHOW VARIABLES LIKE "innodb_buffer_pool_size";'

# Reduce the largest consumer by 10-20% and re-measure free memory

Clear swap without rebooting to confirm which process had pages swapped

Swapping swap off and back on forces all swap pages back into RAM — if RAM is available, this succeeds and the swapped pages return. Use this to clear swap after reducing process memory, then re-examine which processes previously had swap usage.

# Check current swap state
free -h
swapon --show

# Attempt to clear swap (will fail if not enough free RAM)
swapoff -a && swapon -a

# If swapoff fails with 'Cannot allocate memory':
# Not enough free RAM to bring back all swapped pages
# Reduce process memory first (restart PHP-FPM to recycle workers)
systemctl reload php8.1-fpm
sleep 60
swapoff -a && swapon -a

# After successful swapoff/swapon:
# Swap should be 0 and free RAM should be roughly what was swapped + prior free
free -h

Performance Impact: Active Swap I/O Degrades All Processes Simultaneously

Practical Examples

Set vm.swappiness permanently and verify

# Check current value
cat /proc/sys/vm/swappiness

# Set temporarily
sysctl -w vm.swappiness=10

# Persist
cat >> /etc/sysctl.d/99-webserver.conf << 'EOF'
# Prefer file cache reclamation over process swapping
vm.swappiness = 10
# Reduce inode/dentry cache pressure slightly
vm.vfs_cache_pressure = 100
EOF

sysctl -p /etc/sysctl.d/99-webserver.conf

# Verify
cat /proc/sys/vm/swappiness
# Should output: 10

The sysctl.d directory is the correct location for persistent sysctl settings on systemd-based distributions. Settings in /etc/sysctl.d/ are applied at boot before network services start.

Monitor swap I/O rate continuously during a traffic peak

# Watch vmstat every 2 seconds, highlight active swap
watch -n 2 'vmstat 1 3 | awk "NR==1{print} NR==2{print} NR>2{if(\$7>0 || \$8>0) printf \"\\033[31m%s\\033[0m\\n\", \$0; else print}"'

# Alternatively use dstat for a more readable output:
apt install dstat
dstat --swap --disk --mem 2
# Shows real-time swap in/out, disk reads/writes, and free/used/cached memory
# in one view — much easier to correlate swap I/O with disk I/O

dstat is more readable than vmstat for correlating swap I/O with disk I/O. Install it and run during the next reported performance issue to capture the exact conditions.

Troubleshooting Common Issues

Summary

Monitor swap I/O with vmstat — si and so columns show if swapping is actively happening. Non-zero si/so indicates RAM shortage causing performance degradation. Set vm.swappiness=10 to delay swap use until RAM is genuinely exhausted. Identify the largest RAM consumers with ps and reduce their allocation (MySQL buffer pool, PHP-FPM pool size, Apache MaxRequestWorkers). The root fix for chronic swapping is always reducing total process RSS below available RAM.

• Watch si and so columns in vmstat 2 — non-zero values mean active swap I/O is degrading all processes; non-zero swpd with zero si/so is harmless.
• Set vm.swappiness=10 in /etc/sysctl.d/99-webserver.conf — tells the kernel to strongly prefer file cache reclamation over process swapping; default 60 is too aggressive for web server workloads.
• Active swapping always indicates RAM exhaustion — reduce the largest RAM consumer (MySQL innodb_buffer_pool_size, PHP-FPM pm.max_children, or Apache MaxRequestWorkers) until free RAM stays above 200MB under peak load.