<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Rdma on Le Site de François</title><link>https://lesitedefrancois.be/en/tags/rdma/</link><description>Recent content in Rdma on Le Site de François</description><generator>Hugo -- gohugo.io</generator><language>en</language><copyright>© 2026 François</copyright><atom:link href="https://lesitedefrancois.be/en/tags/rdma/index.xml" rel="self" type="application/rss+xml"/><item><title>InfiniBand</title><link>https://lesitedefrancois.be/en/ai/infiniband/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://lesitedefrancois.be/en/ai/infiniband/</guid><description>&lt;p&gt;&lt;strong&gt;InfiniBand&lt;/strong&gt; is a high-performance &lt;strong&gt;network fabric&lt;/strong&gt; designed for datacenter and HPC clusters, natively supporting &lt;strong&gt;RDMA&lt;/strong&gt; (Remote Direct Memory Access) with low latency, high bandwidth, and features such as adaptive routing and congestion control at the link layer. Its objective in &lt;strong&gt;AI&lt;/strong&gt; is to connect many &lt;strong&gt;GPU&lt;/strong&gt; servers so &lt;strong&gt;distributed training&lt;/strong&gt; (gradient all-reduce via &lt;strong&gt;NCCL&lt;/strong&gt;) and multi-node &lt;strong&gt;inference&lt;/strong&gt; (tensor parallel, &lt;strong&gt;llm-d&lt;/strong&gt; prefill/decode &lt;strong&gt;KV&lt;/strong&gt; transfer) are not limited by TCP overhead on a &lt;strong&gt;CPU&lt;/strong&gt;. InfiniBand NICs (e.g. NVIDIA ConnectX) present verbs APIs; subnets are managed with an &lt;strong&gt;Subnet Manager&lt;/strong&gt; and partitioned for multi-tenant isolation.&lt;/p&gt;</description></item><item><title>RDMA (Remote Direct Memory Access)</title><link>https://lesitedefrancois.be/en/ai/rdma/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://lesitedefrancois.be/en/ai/rdma/</guid><description>&lt;p&gt;&lt;strong&gt;RDMA (Remote Direct Memory Access)&lt;/strong&gt; allows a network adapter to transfer data between the memory of two machines with &lt;strong&gt;little CPU overhead&lt;/strong&gt;, low latency, and often &lt;strong&gt;kernel bypass&lt;/strong&gt; (userspace stacks such as &lt;strong&gt;verbs&lt;/strong&gt; on InfiniBand or RoCE). Its objective in AI infrastructure is to keep &lt;strong&gt;GPUs&lt;/strong&gt; fed and synchronized: &lt;strong&gt;distributed training&lt;/strong&gt; exchanges gradients quickly, &lt;strong&gt;disaggregated inference&lt;/strong&gt; (&lt;strong&gt;llm-d&lt;/strong&gt;) moves &lt;strong&gt;KV cache&lt;/strong&gt; blocks between prefill and decode nodes, and &lt;strong&gt;NVMe-oF&lt;/strong&gt; storage delivers checkpoints without the host spending cycles copying every byte. &lt;strong&gt;DPUs&lt;/strong&gt; and &lt;strong&gt;SmartNICs&lt;/strong&gt; also use RDMA paths for storage and east-west traffic while the host CPU runs models.&lt;/p&gt;</description></item><item><title>RoCE (RDMA over Converged Ethernet)</title><link>https://lesitedefrancois.be/en/ai/roce/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://lesitedefrancois.be/en/ai/roce/</guid><description>&lt;p&gt;&lt;strong&gt;RoCE (RDMA over Converged Ethernet)&lt;/strong&gt; implements &lt;strong&gt;RDMA&lt;/strong&gt; semantics on &lt;strong&gt;Ethernet&lt;/strong&gt; (RoCEv2 uses UDP/IP), so NICs can perform remote memory access with low &lt;strong&gt;CPU&lt;/strong&gt; utilization over the same physical switches many enterprises already operate. Its objective is to deliver InfiniBand-like &lt;strong&gt;GPU&lt;/strong&gt; communication economics—fast &lt;strong&gt;NCCL&lt;/strong&gt; all-reduces, &lt;strong&gt;NVMe-oF&lt;/strong&gt;, &lt;strong&gt;llm-d&lt;/strong&gt; KV moves—without maintaining a separate InfiniBand fabric. RoCE requires &lt;strong&gt;lossless Ethernet&lt;/strong&gt; behavior: Priority Flow Control (&lt;strong&gt;PFC&lt;/strong&gt;), Explicit Congestion Notification (&lt;strong&gt;ECN&lt;/strong&gt;), buffer tuning, and often dedicated traffic classes so RDMA traffic is not dropped under burst load.&lt;/p&gt;</description></item></channel></rss>