MCP CPU Spike

[derekhiggins]

MCP CPU Spike

Summary

When making MCP calls through the responses API, the llamastack server process CPU usage spikes to 100% and remains there indefinitely, even after the request completes.

Environment

• LLamaStack version: main branch
• Python version: 3.12

Steps to Reproduce

1. Start llamastack server:
2. Verify CPU usage is idle (0-1%)
3. Make an MCP call via responses API:
4. Monitor CPU usage with top

Expected Behavior

CPU usage should return to idle levels (0-1%) after the MCP request completes.

Actual Behavior

CPU usage spikes to 100% and stays there indefinitely:

# Before MCP call - idle
3176764 derekh    20   0 4196036 442736 137316 S   0.0   1.4   0:06.55 llama stack run

# After MCP call - stuck at 100%
3176764 derekh    20   0 4422628 448496 137444 R  99.7   1.4   0:34.23 llama stack run

Root Cause

It looks like the issue is caused by the MCP session caching mechanism (MCPSessionManager) that
was added to optimize performance by avoiding redundant tools/list calls (fix for #4452).

[derekhiggins]

cc @jwm4

[jwm4]

Thx, @derekhiggins . I will try to figure out what's wrong here.

[jwm4]

@derekhiggins , PR for a fix is up at #4758 . I checked in top with this fix and I didn't see any problems with the CPU, but you might want to try it out too.

[mattf]

note: using a tool is not necessary to reproduce this cpu consumption. it's sufficient for a client to mention an mcp server.

[othman1387]

MCP CPU Spike

Summary

When making MCP calls through the responses API, the llamastack server process CPU usage spikes to 100% and remains there indefinitely, even after the request completes.

Environment

• LLamaStack version: main branch
• Python version: 3.12

Steps to Reproduce

1. Start llamastack server:
2. Verify CPU usage is idle (0-1%)
3. Make an MCP call via responses API:
4. Monitor CPU usage with top

Expected Behavior

CPU usage should return to idle levels (0-1%) after the MCP request completes.

Actual Behavior

CPU usage spikes to 100% and stays there indefinitely:

# Before MCP call - idle
3176764 derekh    20   0 4196036 442736 137316 S   0.0   1.4   0:06.55 llama stack run

# After MCP call - stuck at 100%
3176764 derekh    20   0 4422628 448496 137444 R  99.7   1.4   0:34.23 llama stack run

Root Cause

It looks like the issue is caused by the MCP session caching mechanism (MCPSessionManager) that
was added to optimize performance by avoiding redundant tools/list calls (fix for #4452).