3.8 Addressing Challenges in Today's Datacenter Systems' Design

Since its inception, CloudSuite (cloudsuite.ch) has emerged as a popular suite of benchmarks both in industry and among academics for the performance evaluation of cloud services. The EuroCloud Server project blueprinted key optimizations in server SoCs based on the salient features of CloudSuite benchmarks that lead to an order of magnitude improvement in efficiency while preserving QoS. ARM-based server products (e.g., Cavium ThunderX) have now emerged following these guidelines and showcasing the improved efficiency. CloudSuite 3.0 is a major enhancement over prior releases both in benchmarks and infrastructure. It includes benchmarks that represent massive data manipulation with tight latency constraints such as in-memory data analytics using Apache Spark, a new real-time video streaming benchmark following today's most popular video-sharing website setups, and a new web serving benchmark mirroring today's multi-tier web server software stacks. To ease the deployment of CloudSuite into private and public cloud systems, the benchmarks are integrated into the Docker software container system and Google's PerfKit Benchmarker. Docker wraps each benchmark into a self-contained software package, guaranteeing the same execution regardless of the environment, while PerfKit automates the process of benchmarking cloud server systems with CloudSuite. CloudSuite 3.0 is supported to run both on real hardware and on our QEMU-based computer architecture simulation framework.

FaRM is an in-memory, transactional database that runs distributed across a cluster of Windows Servers that are connected by a high-speed Remote Direct Memory Access (RDMA) network. Data in FaRM are stored in DRAM and exposed directly to the L2 network by the server's RDMA network adapters, so that other members of the FaRM cluster can access the data with great efficiency. RDMA enables a network adapter to directly access the memory of another server in the same Ethernet network bypassing the operating system in both servers. This enables low-latency and high-bandwidth data access across the entire cluster. However, RDMA provides no security: the data are also accessible to every other server attached to the same Ethernet network, and message transfers are vulnerable to replay and modification. We present our work to protect data in FaRM using a bump-in-the-wire firewall for RDMA. Based upon the FPGA cards widely deployed in Windows Servers within Microsoft, the firewall exists as a barrier between a FaRM server's RDMA adapter and the local Ethernet switch. It prevents packets from outside the FaRM cluster from ever reaching the server's RDMA adapter, and it protects RDMA packets between members of the FaRM cluster by encapsulating them in DTLS tunnels. We show that implementing a similar level of protection in software can be prohibitively expensive.