HydraRPC: RPC in the CXL Era
summary
HydraRPC implements an RPC stack that mounts CXL-attached HDM, exposes 64-bit message-queue entries and preallocated data slots, and removes network round trips, memcpy and serialization by passing references into shared memory. It uses non-cachable CXL HDM with monitor/mwait notification to avoid busy polling, a user-space allocator and sliding-window flow control, and runs on real CXL 1.1 hardware. Measured latency is ~1.47 ms average with 99th-percentile 4x improvement and up to ≈100x speedups for 16B–2MB payloads versus RDMA/MRPC baselines.
AI-written summary · about
chapters (11)
- 0:00HydraRPC and CXL intro
- 2:08CXL shared memory infrastructure
- 4:13Traditional RPC/RDMA limits
- 6:03HydraRPC architecture design
- 7:31Memory management in HydraRPC
- 9:00Request/response workflow
- 11:13Bypassing CPU cache
- 12:37Notification mechanisms
- 13:42Memory allocation & sliding window
- 15:30Performance results
- 17:24Summary & Q&A
related talks
17mCeph Messenger Situation and Optimization
17mDRack: A CXL-Disaggregated Rack Architecture to Boost Inter-Rack Communication
24mEnhancing Homa Linux for Efficient RPC Transportation
14mODRP: On-Demand Remote Paging with Programmable RDMA
24mDatacenter RPCs can be General and Fast
17mScalable Distributed Massive MIMO Baseband Processing
20mKRCORE: A Microsecond-scale RDMA Control Plane for Elastic Computing
18mRDMA is Turing complete, we just did not know it yet!