Hardware transports such as RoCE deliver high performance with minimal host CPU, but are best suited to special-purpose deployments that limit their use, e.g., backend networks or Ethernet with Priority Flow Control (PFC). We introduce Falcon, the first hardware transport that supports multiple Upper Layer Protocols (ULPs) and heterogeneous application workloads in general-purpose Ethernet datacenter environments (with losses and without special switch support). Key design elements include: delay-based congestion control with multipath load balancing; a layered design with a simple request-response transaction interface for multi-ULP support; hardware-based retransmissions and error-handling for scalability; and a programmable engine for flexibility. The first Falcon hardware implementation delivers a peak performance of 200 Gbps, 120 Mops/sec, with near-optimal operation completion times that are up to 8× lower than CX-7 RoCE under network congestion, and up to 65% higher goodput under lossy conditions.

We present evidence and characterization of host congestion in production clusters: adoption of high-bandwidth access links leading to emergence of bottlenecks within the host interconnect (NIC-to-CPU data path). We demonstrate that contention on existing IO memory management units and/or the memory subsystem can significantly reduce the available NIC-to-CPU bandwidth, resulting in hundreds of microseconds of queueing delays and eventual packet drops at hosts (even when running a state-of-the-art congestion control protocol that accounts for CPU-induced host congestion). We also discuss implications of host interconnect congestion to design of future host architecture, network stacks and network protocols.

We report on experiences with Swift congestion control in Google datacenters. Swift targets an end-to-end delay by using AIMD control, with pacing under extreme congestion. With accurate RTT measurement and care in reasoning about delay targets, we find this design is a foundation for excellent performance when network distances are well-known. Importantly, its simplicity helps us to meet operational challenges. Delay is easy to decompose into fabric and host components to separate concerns, and effortless to deploy and maintain as a congestion signal while the datacenter evolves. In large-scale testbed experiments, Swift delivers a tail latency of <50μs for short RPCs, with near-zero packet drops, while sustaining ~100Gbps throughput per server. This is a tail of <3x the minimal latency at a load close to 100%. In production use in many different clusters, Swift achieves consistently low tail completion times for short RPCs, while providing high throughput for long RPCs. It has loss rates that are at least 10x lower than a DCTCP protocol, and handles O(10k) incasts that sharply degrade with DCTCP.

Remote Direct Memory Access (RDMA) plays a key role in supporting performance-hungry datacenter applications. However, existing RDMA technologies are ill-suited to multi-tenant datacenters, where applications run at massive scales, tenants require isolation and security, and the workload mix changes over time. Our experiences seeking to operationalize RDMA at scale indicate that these ills are rooted in standard RDMA's basic design attributes: connectionorientedness and complex policies baked into hardware.

Homa is a new transport protocol for datacenter networks. It provides exceptionally low latency, especially for workloads with a high volume of very short messages, and it also supports large messages and high network utilization. Homa uses in-network priority queues to ensure low latency for short messages; priority allocation is managed dynamically by each receiver and integrated with a receiver-driven flow control mechanism. Homa also uses controlled overcommitment of receiver downlinks to ensure efficient bandwidth utilization at high load. Our implementation of Homa delivers 99th percentile round-trip times less than 15 μs for short messages on a 10 Gbps network running at 80% load. These latencies are almost 100x lower than the best published measurements of an implementation. In simulations, Homa's latency is roughly equal to pFabric and significantly better than pHost, PIAS, and NDP for almost all message sizes and workloads. Homa can also sustain higher network loads than pFabric, pHost, or PIAS.

RAMCloud is a storage system that provides low-latency access to large-scale datasets. To achieve low latency, RAMCloud stores all data in DRAM at all times. To support large capacities (1PB or more), it aggregates the memories of thousands of servers into a single coherent key-value store. RAMCloud ensures the durability of DRAM-based data by keeping backup copies on secondary storage. It uses a uniform log-structured mechanism to manage both DRAM and secondary storage, which results in high performance and efficient memory usage. RAMCloud uses a polling-based approach to communication, bypassing the kernel to communicate directly with NICs; with this approach, client applications can read small objects from any RAMCloud storage server in less than 5μs, durable writes of small objects take about 13.5μs. RAMCloud does not keep multiple copies of data online; instead, it provides high availability by recovering from crashes very quickly (1 to 2 seconds). RAMCloud’s crash recovery mechanism harnesses the resources of the entire cluster working concurrently so that recovery performance scales with cluster size.