Search for: All records

Luiz Barroso started his career at Digital Equipment Corporation, investigating workload-optimized multiprocessor server architectures marketed to enterprises in the 1990s. These high-margin, low-volume products lost their market to more cost-effective enterprise servers built from high-volume desktop CPUs riding Moore’s law. The enterprise market has slowly transitioned to the cloud, where desktop PCs have formed the backbone of computing in data centers since the early 2000s to minimize cost and maximize the return on investment. Moving forward, with the absence of Moore’s law, future servers require a clean-slate, cross-stack design to scale in compute, communication, and storage capacity while reducing operational, capital, and environmental costs. 

Server applications exhibit a high degree of code repetition because they handle many similar requests. In turn, repeated execution of the same code, often with identical inputs, highlights an inefficiency in the execution of server software and suggests memoization as a way to improve performance. Memoization has been extensively explored in software, and several hardware- and hardware-assisted memoization schemes have been proposed in the literature. However, these works targeted memoization of mathematical or algorithmic processing, whereas server applications call for a different approach. We observe that the opportunity for memoization in servers arises not from eliminating the repetition of complex computation, but from eliminating the repetition of software orchestration code. This work studies hardware memoization in servers, ultimately focusing on one pattern, instruction sequences starting with indirect jumps.We explore how an out-of-order pipeline can be extended to support memoization of these instruction sequences, demonstrating the potential of hardware memoization for servers. Using 26 applications to make our case (3 CloudSuite workloads and 23 vSwarm serverless functions), we show how targeting just this one pattern of instruction sequences can memoize over 10% (up to 15.6%) of the dynamically executed instructions in these server applications. 

Modern last-level caches are partitioned into slices that are spread across the chip, giving rise to varying access latencies dictated by the physical location of the accessing core and the cache slice being accessed. Although, prior work has shown that dynamically determining the best location for blocks within such Non-Uniform Cache Access architectures can provide significant performance benefits, current hardware does not implement this functionality. Instead, modern processors hash blocks across the LLC slices, obscuring the non-uniform architecture of the underlying cache and forfeiting the performance benefits of placing data in the nearest cache slices. Moreover, while prior work advocated improving performance by delegating control over block placement to the operating system at page granularity, modern processor hardware thwarts these approaches by hashing cache slice selection at cache block granularity. In this work, we make two observations that enable us to improve software performance on modern NUCA architectures. First, we find that software can undo the hashing performed by hardware and efficiently manage data placement at cache block granularity. Second, that the complexity of fine-grained data placement can be hidden from the developer by embedding it in the dynamic memory allocator. Leveraging these observations, we design a new specialized memory allocator, NUCAlloc, suitable for use with C++ containers such as std::map and std::set. NUCAlloc handles the complexity of NUCA-aware block placement, improving the performance of containers by placing their data into the nearest LLC slices. We demonstrate that our NUCAlloc prototype consistently outperforms std::allocator and jemalloc for LLC-resident containers, improving performance by up to 20% in both single-threaded and multi-threaded software.