peercache 0.6.5

Peer-to-peer RDMA zero-copy L3 KV-cache backend for SGLang HiCache

PeerCache

English · 简体中文

A lightweight, peer-to-peer L3 storage backend for SGLang HiCache, built for PD-disaggregated (prefill/decode) inference: prefill workers publish KV pages, decode workers read them back over RDMA with zero CPU copies.

Docs: https://flymysql.github.io/PeerCache/

PeerCache gives you Mooncake-style RDMA zero-copy KV-cache sharing across nodes, but without the centralized master + metadata services. Instead it uses:

• Embedded service discovery — no separate meta process. One node (chosen by discovery_addr) auto-hosts the discovery service in-process; nodes register their endpoint, heartbeat, and pull the live membership list.
• A consistent-hash distributed directory (DHT) — the mapping key -> {data_node, remote_addr, rkey, length} is sharded across all nodes by hashing the key. There is no central metadata store.
• Data stays local on write — set() copies the page into a node-local published pool (a host memcpy, no network, no master) and pushes only a tiny location record to the directory.
• One-sided RDMA READ on read — get() looks up the directory, then issues a zero-copy IBV_WR_RDMA_READ straight into SGLang's registered host buffer.
• Concurrent multi-threaded I/O — a per-peer channel pool (each an RC QP with its own completion queue) lets reader/writer threads run with no shared-CQ contention; the control plane parallelises directory lookups too.
• Disk persistence tier (L4) — pages evicted from memory spill to disk (default /data/peercache/, 100GB) and are promoted back into the pool on a later read (locally or by a remote reader).
• Built-in monitoring — Prometheus /metrics + an embedded HTML dashboard (default port 31997): hit rate, throughput, latency p50/p99, mem/disk usage.

Architecture

PeerCache splits into a C++ data plane (raw RDMA verbs) and a Python control plane (discovery, directory, pool). The meta node only tracks membership; the key -> location map is sharded across every node by consistent hashing, and the KV bytes never leave the node that produced them until a peer reads them directly.

flowchart LR
    subgraph A["Node 0 — prefill / producer"]
        S0["PeerCacheStore"]
        P0[("Published pool MR (LRU)")]
        D0["Directory shard"]
    end
    subgraph B["Node 1 — decode / consumer"]
        S1["PeerCacheStore"]
        R1[("Receive MR = SGLang kv_buffer")]
        D1["Directory shard"]
    end
    M(["Embedded discovery / membership"])

    S0 -- "set(): local memcpy" --> P0
    S0 -- "PUT key→loc (hash)" --> D1
    S1 -- "GET key (hash)" --> D0
    S1 == "one-sided RDMA READ (zero copy)" ==> P0
    P0 -. "lands directly in" .-> R1
    S0 -. "register / heartbeat" .-> M
    S1 -. "register / heartbeat" .-> M

write:  set() ── local memcpy ──> published pool MR
                └── PUT key->{node,addr,rkey,len} ──> directory shard (hash(key))
read:   get() ── GET key ──> directory shard ──> {node,addr,rkey,len}
                └── one-sided RDMA READ ──> local host buffer (zero copy)

• C++ data plane (cpp/): raw libibverbs + librdmacm. RC QPs, one-sided READ/WRITE, per-peer channel pool with private CQs for concurrency. Exposed to Python via pybind11 as the _peercache module.
• Python control plane (python/peercache/): TCP RPC, service discovery, consistent-hash ring, distributed directory, and the published-pool with LRU.
• TCP fallback transport: a pure-Python transport that mirrors the RDMA API so the design can be validated end-to-end on machines without RDMA hardware.

Two-MR model (correctness)

SGLang's host KV buffer is the L2 tier and is evicted/overwritten by HiCache, so we cannot register its address into the directory directly (dangling reference). Each node therefore registers two memory regions:

1. Receive MR = mem_pool_host.kv_buffer — destination of one-sided READ on get.
2. Published pool MR = a backend-owned host pool with LRU — source of READ on remote nodes. set memcpys the page into this pool (node-local, no network) and publishes its addr+rkey+len to the directory. Eviction from the pool deletes the corresponding directory entry, so a published address stays valid until evicted.

See the Architecture docs for the full design.

Why simpler than Mooncake?

	Mooncake	PeerCache
metadata	central master + metadata service	sharded directory (consistent hash)
data placement	dedicated managed pool	stays on producing node
coordination	master allocates / tracks objects	only service discovery on meta node
transfer	RDMA zero-copy	RDMA zero-copy (one-sided READ)

Install

From PyPI (recommended)

pip install peercache

This builds the C++ data plane from source, so the target host needs a C++17 toolchain, CMake ≥ 3.18, and the RDMA dev headers (libibverbs / librdmacm, e.g. rdma-core or Mellanox OFED). If those headers are absent, the build automatically falls back to a stub module and the pure-Python TCP transport.

To force the no-RDMA build explicitly (control plane + TCP fallback only, e.g. on a laptop or in CI):

pip install peercache --config-settings=cmake.define.PEERCACHE_NO_RDMA=ON

From source

git clone https://github.com/flymysql/PeerCache.git
cd PeerCache
pip install .                 # or: pip install -e ".[test]"

Run with SGLang

The meta service is embedded — there is no separate meta process. Point discovery_addr at one node's IP on every node; the node whose IP matches auto-starts the discovery service in-process.

# On every SGLang node, set discovery_addr to the SAME node's IP (say node-0).
# node-0 detects the IP is itself and hosts the embedded meta automatically.
python -m sglang.launch_server --enable-hierarchical-cache \
  --hicache-storage-backend dynamic \
  --hicache-storage-backend-extra-config \
  '{"backend_name":"peercache","module_path":"peercache.store","class_name":"PeerCacheStore","discovery_addr":"NODE0_IP:31998","protocol":"rdma","device_name":"mlx5_0","global_segment_size":"4gb"}'

(Optionally, you can still run a standalone meta with peercache-meta --bind 0.0.0.0:31998 if you prefer a dedicated discovery host.)

See examples/sglang_launch.md for details.

Benchmarks

A systematic benchmark suite ships inside the package and is exposed as a single console command (no repo clone, no PYTHONPATH). It drives PeerCache's HiCacheStorage interface exactly as SGLang HiCache does (PD-disaggregated batch_set_v1 / batch_exists / batch_get_v1) and reports throughput (pages/s, tokens/s, GB/s) and latency tail (p50/p95/p99/p999/max) across a sweep of thread models, including the full-load saturation/peak throughput.

Measured baseline (cross-host RDMA, GET, MLA)

On 2× AMD EPYC 9K84 + 8× ConnectX-7 (RoCEv2, MTU 4096, MLNX_OFED 5.8):

scenario	GET throughput
single NIC, PeerCache	46.0 GB/s (368 Gbps) — ~94% of bare ib_read_bw (49.0 GB/s)
single process, 8 rails (1 MiB pages)	147.6 GB/s (1.18 Tbps)
full machine, 8 NICs, multi-process	413.1 GB/s (≈ 3.3 Tbps)

Methodology, charts, and reproduce commands: Performance baseline.

pip install peercache

# RDMA hardware (publishable numbers)
peercache-bench suite --device-name mlx5_0 --layout mla --page-size 131072 \
    --batch-size 32 --concurrencies 1,2,4,8,16,32,64 --duration 10 --tag rdma

peercache-bench subcommands: latency, throughput, saturation, suite (SGLang-HiCache), plus micro, mooncake, compare.

RDMA-first. PeerCache is built on RDMA one-sided READ; publishable figures must be measured on RDMA hardware. The TCP fallback is for functional smoke testing only and must not be quoted. See the bench README and the Benchmarks docs for the methodology, thread models, metric definitions, and reproduction recipe.

Test

pip install -e ".[test]"
pytest -q

Maintainer setup (one-time)

• GitHub Pages: Settings → Pages → Build and deployment → Source = GitHub Actions. The Docs workflow then publishes to https://flymysql.github.io/PeerCache/ on every push to main.
• PyPI Trusted Publishing: on the PyPI peercache project, add a GitHub publisher (owner flymysql, repo PeerCache, workflow release.yml, environment pypi). Tagging vX.Y.Z then builds the sdist, attaches it to a GitHub Release, and publishes to PyPI. Until configured, the PyPI step is non-blocking and the GitHub Release still ships the package.