Blog

LMCache · August 2026 Optimizing LMCache's prefix lookup for one-sided RDMA How LMCache's existence check works, why it does it (prefix caching), the batched contains it already uses, and how a one-sided RDMA backend folds the read address into that check and releases the GIL. Read → Guide · August 2026 Deploy Rasa in your own AWS account with the CDK What the stack creates, every context parameter, and the controls the console gives you, from adding a data node to draining and rotating arenas. Read → Benchmark · August 2026 Connections, queue depth, and batch reads: tuning a one-sided RDMA cache How many connections each size needed to saturate EFA (32 for 4 KiB, one from 4 MiB up), why queue depth is the cheaper knob, the security-group rule that silently drops all EFA traffic, and the write path that turned out to be CPU-bound. Read → Insight · August 2026 How a one-sided RDMA cache stays correct without trusting the reader No server runs on the read path. One decision makes that safe: values are immutable. Write-once removes torn and stale reads, leaving three guards: a signed descriptor, a checksum over every read, and rkey rotation as a hardware fence on the one moment memory is reused. Why the worst case is always a recompute, never a wrong value. Read → Design · July 2026 When the cache server writes the answer into you One-sided reads are a pull, but a value on NVMe can't be pulled because the NIC can't fi_read block storage. Flip the direction: the client hands the node a write-key and the node RDMA-writes the value straight into the client's buffer. One round trip, the reply is the completion. Read → Deep dive · July 2026 How Rasa writes: allocate, RDMA-write, commit RDMA writes move bytes quickly. The correctness comes from allocation grants, commit ordering, signed descriptors, and not publishing half-written memory. Read → Deep dive · July 2026 How Rasa reads: descriptors, bulk reads, and the cost of moving memory One-sided reads are simple once the client has a descriptor. The work is in descriptor batching, keeping the client map small, and keeping remote memory stable. Read → Insight · June 2026 Why mutability is hard with RDMA One-sided reads are fast because nobody is home on the read path, which also makes changing a value in place hard: torn reads, no read-side validation, and memory you cannot safely reclaim. Why immutability is the usual answer. Read → Architecture · June 2026 Rasa architecture in one page A system map: client SDK, Valkey control plane, data nodes, registered arenas, EFA/libfabric, and the rule we used to decide what belongs where. Read → Deep dive · June 2026 SRD: the cloud-native transport under EFA Why AWS replaced TCP and InfiniBand inside the cloud with SRD: packet spraying across many paths, out-of-order delivery, microsecond hardware retransmission in the Nitro card, and where it powers EFA, ENA Express, and EBS. Read → Primer · June 2026 RDMA on AWS: what it is, how it works, and what makes EFA different Kernel bypass and one-sided reads, why RDMA is hard, how it differs from a normal key/value GET, what SRD changes, and why you cannot tell from an instance name whether it supports a one-sided read. Read →