Dibs

Abstract

Data centers must support a range of workloads with differing demands. Although existing approaches handle routine traffic smoothly, intense hotspots-even if ephemeralcause excessive packet loss and severely degrade performance. This loss occurs even though congestion is typically highly localized, with spare buffer capacity at nearby switches. In this paper, we argue that switches should share buffer capacity to effectively handle this spot congestion without the monetary hit of deploying large buffers at individual switches. Specifically, we present detour-induced buffer sharing (DIBS), a mechanism that achieves a near lossless network without requiring additional buffers at individual switches. Using DIBS, a congested switch detours packets randomly to neighboring switches to avoid dropping the packets. We implement DIBS in hardware, on software routers in a testbed, and in simulation, and we demonstrate that it reduces the 99th percentile of delay-sensitive query completion time by up to 85%, with very little impact on other traffic.

Dart: Divide and Specialize for Fast Response to Congestion in RDMA-Based Datacenter Networks

“…As such, deflection is relatively straightforward to use when either the application does not require ordered packet delivery or there is a reassembly layer that reorders received packets to be delivered in the correct order (e.g., TCP). Indeed, in addition to being well-explored in other contexts, such packet-by-packet deflection has also been proposed for congestion avoidance in data centers (DIBS [52]). In contrast, for RDMA networks, there is no software stack to reassemble out-of-order packets of a message/flow.…”

“…Presto [24] splits large flows into equal-sized flowcells and uses a central scheduler to balance the load. DeTail [18], Random Packet Spraying [13], and DIBS [52] balance load at the finer granularity of packets. Recent schemes [21,53] improve load balancing but they also require reordering at the receiver.…”

“…For the latter where the network capacity is under pressure (e.g., due to dynamic network load spikes), Dart falls back on DCQCN's throttling which may be unavoidable. Load balancing [2,13,18,24,30,39,42,52] can alleviate localized in-network congestion but not receiver congestion, and usually reorders packets which is not supported by RDMA.…”

“…Because RDMA does not support packet reordering, Dart provides hardware support in the switch to keep a flow's packets in order. While deflection [5] is well known, our contribution is in-order flow deflection (IOFD) unlike previous load-balancing schemes including DIBS [52]. As a congestion response, deflection is much lighter-weight and quicker (well under one RTT) than rate-cutting using iterative convergence and does not affect the sending rates.…”