Mooncake Fragmentation-Aware Allocation Strategy

Competition: The 8th CCF Open Source Innovation Competition, Open Source Task Challenge Track.

Task: Mooncake KVCache storage design and performance optimization.

Work Summary

This submission adds a new Mooncake Store allocation strategy named fragmentation_aware. The strategy improves allocation decisions for mixed-size KV cache workloads by considering the largest contiguous free region in each candidate segment, not only the aggregate free-space ratio.

The feature is designed to reduce failed allocation attempts in long-running clusters where memory becomes fragmented. It keeps Mooncake’s existing best-effort replica semantics and uses a bounded sampled candidate set to avoid full-cluster scans.

Key Deliverables

• New AllocationStrategyType::FRAGMENTATION_AWARE.
• New FragmentationAwareAllocationStrategy.
• Master startup support via --allocation_strategy=fragmentation_aware.
• Unit test covering a deterministic fragmentation scenario.
• Allocation benchmark matrix support for the new strategy.
• Updated design, deployment, and quick-start documentation.
• TCP-only / old-libibverbs build compatibility fixes needed for local validation on Ubuntu 20.04.

Repository

Main working repository:

C:\CCFOpenSource\Repos\Mooncake-github

Working branch:

ccf-fragmentation-aware-allocation

Usage

Start Mooncake master with:

mooncake_master --allocation_strategy=fragmentation_aware

For HTTP metadata quick start:

mooncake_master \
  --allocation_strategy=fragmentation_aware \
  --enable_http_metadata_server=true \
  --http_metadata_server_port=8080

Status

Implementation patch is prepared at:

C:\CCFOpenSource\Submission\mooncake_fragmentation_aware.patch

Local verification results are recorded in C:\CCFOpenSource\Submission\testing.md and C:\CCFOpenSource\Submission\logs.

Primary local validation:

• allocation_strategy_light_test.log: passed.
• fragmentation_aware_sim.log: passed.
• Full upstream Store target was attempted but did not finish in the local WSL/GCC10/Windows-mounted workspace within the available resource window.