Why Does Java Service Load Spike on Startup? Uncover C2 CompilerThread Impact & Fixes

Problem Description

During the deployment or restart of an online service (Jetty8 as the server), some machines experience a load spike up to 70 that lasts for about five minutes, and the response‑time curve mirrors the load curve.

Investigation Method

Monitor resource usage during publishing.

1) Use top -H -p to find high‑CPU threads; threads 2129 and 2130 show high usage.

2) Use jstack to dump stack information, convert thread IDs to hexadecimal (2129 → 851, 2130 → 852) and identify them as compiler threads (see Table 1).

"C2 CompilerThread1" daemon prio=10 tid=0x00007fce48125800 nid=0x852 waiting on condition [0x0000000000000000]
java.lang.Thread.State: RUNNABLE
Locked ownable synchronizers:
- None
"C2 CompilerThread0" daemon prio=10 tid=0x00007fce48123000 nid=0x851 waiting on condition [0x0000000000000000]
java.lang.Thread.State: RUNNABLE
Locked ownable synchronizers:
- None

Explanation of the Phenomenon

At startup, Java runs in interpreted mode, which is slow and causes high load and response times. As traffic flows, frequently executed methods are compiled by the C2 CompilerThread, increasing its CPU usage. Once compilation finishes, the thread’s CPU usage drops, performance improves, and load decreases.

The root cause is the prolonged compilation‑optimization phase, which creates the observed jitter.

Solutions

1) Warm‑up

Complete compilation before receiving real traffic. Two approaches:

Programmatic warm‑up: after startup, invoke hot code paths so they are compiled (verify with -XX:+PrintCompilation).

Traffic replay: use tools like tcpcopy to replay production traffic for warm‑up.

2) Increase compiler threads

Use the -XX:CICompilerCount flag to raise the number of compiler threads (default 2) to, for example, 4.

3) Use multi‑level compilation

HotSpot offers three compilation modes:

Client mode – lightweight, fast trigger, lower optimization.

Server mode – higher optimization, slower trigger (default for the service).

Tiered mode – starts with Client, then switches to Server, combining fast start‑up with high final performance. Enable with -XX:+TieredCompilation (default in Java 8).

Result Analysis

Applying warm‑up (solution 1) together with multi‑threaded compilation (solution 2) and tiered mode (solution 3) reduced load spikes dramatically. After several deployments, most machines stayed within a load range of 2–4, and spikes above 10 were rare. The combined approach shortens the jitter duration and improves overall stability.