Hardware offloaded network stacks are widely adopted in modern datacenters to meet the demand for high throughput, ultra-low latency and low CPU overhead. To fully leverage their exceptional performance, users need to have a deep understanding of their behaviors. Despite many efforts on testing software network stacks, hardware network stacks impose unique challenges to testing tools due to their kernel bypass nature and high performance.

Storage systems in data centers are an important component of large-scale online services. They typically perform replicated transactional operations for high data availability and integrity. Today, however, such operations suffer from high tail latency even with recent kernel bypass and storage optimizations, and thus affect the predictability of end-to-end performance of these services. We observe that the root cause of the problem is the involvement of the CPU, a precious commodity in multi-tenant settings, in the critical path of replicated transactions. In this paper, we present HyperLoop, a new framework that removes CPU from the critical path of replicated transactions in storage systems by offloading them to commodity RDMA NICs, with non-volatile memory as the storage medium. To achieve this, we develop new and general NIC offloading primitives that can perform memory operations on all nodes in a replication group while guaranteeing ACID properties without CPU involvement. We demonstrate that popular storage applications can be easily optimized using our primitives. Our evaluation results with microbenchmarks and application benchmarks show that HyperLoop can reduce 99th percentile latency ≈ 800X with close to 0% CPU consumption on replicas.

Direct network I/O allows network controllers (NICs) to expose multiple instances of themselves, to be used by untrusted software without a trusted intermediary. Direct I/O thus frees researchers from legacy software, fueling studies that innovate in multitenant setups. Such studies, however, overwhelmingly ignore one serious problem: direct memory accesses (DMAs) of NICs disallow page faults, forcing systems to either pin entire address spaces to physical memory and thereby hinder memory utilization, or resort to APIs that pin/unpin memory buffers before/after they are DMAed, which complicates the programming model and hampers performance.

With the tremendous popularity gained by container technology, many applications are being containerized: splitting into numerous containers connected by networks. However, current container networking solutions have either bad performance or poor portability, which undermines the advantages of containerization. In this paper, we propose FreeFlow, a container networking solution which achieves both high performance and good portability. FreeFlow is designed according to the observation that strict isolations are unnecessary among containers trusting each other, and it can significantly boost the communication quality of containers by compromising isolation a little bit. Specifically, we enable containers on the same physical machine to communicate via shared-memory and the ones on different physical machines communicate via high performance networking options, e.g. RDMA and DPDK. Naively wrapping up all the solutions together will result in poor potability of containers and huge complexity in application development. Instead, FreeFlow leverages a network abstraction which supports all common network APIs and a centralized network orchestrator which decides how to deliver data transparently to applications in the containers.

Modern datacenter applications demand high throughput (40Gbps) and ultra-low latency (< 10 μs per hop) from the network, with low CPU overhead. Standard TCP/IP stacks cannot meet these requirements, but Remote Direct Memory Access (RDMA) can. On IP-routed datacenter networks, RDMA is deployed using RoCEv2 protocol, which relies on Priority-based Flow Control (PFC) to enable a drop-free network. However, PFC can lead to poor application performance due to problems like head-of-line blocking and unfairness. To alleviates these problems, we introduce DCQCN, an end-to-end congestion control scheme for RoCEv2. To optimize DCQCN performance, we build a fluid model, and provide guidelines for tuning switch buffer thresholds, and other protocol parameters. Using a 3-tier Clos network testbed, we show that DCQCN dramatically improves throughput and fairness of RoCEv2 RDMA traffic. DCQCN is implemented in Mellanox NICs, and is being deployed in Microsoft's datacenters.

This work addresses the scalability and efficiency of RAM-based storage systems wherein multiple objects must be retrieved per user request. Here, much of the CPU work is per server transaction, not per requested item. Adding servers and spreading the data across them also spreads any given set of requested items across more servers, thereby increasing the total number of server transactions per user request. The resulting poor scalability, dubbed the Multi-get Hole, has been reported in Web 2.0 systems using memcached - a popular memory-based key-value storage system. We present Replicate and Bundle (RnB), a somewhat unintuitive approach: rather than add CPUs, we add memory. Object replicas are mapped “randomly” to servers, and requested objects are bundled, selecting replicas so as to minimize the number of servers accessed per user request and thus the total CPU work per request. We studied RnB via simulation in the context of DRAM-based storage, utilizing micro benchmarks and implemented RnB modules for calibration. Our results show that RnB substantially reduces the number of transactions per request, making operation more efficient. Also, unlike most alternatives, RnB permits flexible growth and relatively easy deployment. Finally, in systems wherein data is replicated for other reasons, RnB is nearly free.