Daniel Sanchez

ZSim: fast and accurate microarchitectural simulation of thousand-core systems

Architectural simulation is time-consuming, and the trend towards hundreds of cores is making sequential simulation even slower. Existing parallel simulation techniques either scale poorly due to excessive synchronization, or sacrifice accuracy by allowing event reordering and using simplistic contention models. As a result, most researchers use sequential simulators and model small-scale systems with 16-32 cores. With 100-core chips already available, developing simulators that scale to thousands of cores is crucial. We present three novel techniques that, together, make thousand-core simulation practical. First, we speed up detailed core models (including OOO cores) with instruction-driven timing models that leverage dynamic binary translation. Second, we introduce bound-weave, a two-phase parallelization technique that scales parallel simulation on multicore hosts efficiently with minimal loss of accuracy. Third, we implement lightweight user-level virtualization to support complex workloads, including multiprogrammed, client-server, and managed-runtime applications, without the need for full-system simulation, sidestepping the lack of scalable OSs and ISAs that support thousands of cores. We use these techniques to build zsim, a fast, scalable, and accurate simulator. On a 16-core host, zsim models a 1024-core chip at speeds of up to 1,500 MIPS using simple cores and up to 300 MIPS using detailed OOO cores, 2-3 orders of magnitude faster than existing parallel simulators. Simulator performance scales well with both the number of modeled cores and the number of host cores. We validate zsim against a real Westmere system on a wide variety of workloads, and find performance and microarchitectural events to be within a narrow range of the real system.

Vantage: scalable and efficient fine-grain cache partitioning

Cache partitioning has a wide range of uses in CMPs, from guaranteeing quality of service and controlled sharing to security-related techniques. However, existing cache partitioning schemes (such as way-partitioning) are limited to coarse-grain allocations, can only support few partitions, and reduce cache associativity, hurting performance. Hence, these techniques can only be applied to CMPs with 2-4 cores, but fail to scale to tens of cores. We present Vantage, a novel cache partitioning technique that overcomes the limitations of existing schemes: caches can have tens of partitions with sizes specified at cache line granularity, while maintaining high associativity and strong isolation among partitions. Vantage leverages cache arrays with good hashing and associativity, which enable soft-pinning a large portion of cache lines. It enforces capacity allocations by controlling the replacement process. Unlike prior schemes, Vantage provides strict isolation guarantees by partitioning most (e.g. 90%) of the cache instead of all of it. Vantage is derived from analytical models, which allow us to provide strong guarantees and bounds on associativity and sizing independent of the number of partitions and their behaviors. It is simple to implement, requiring around 1.5% state overhead and simple changes to the cache controller. We evaluate Vantage using extensive simulations. On a 32-core system, using 350 multi programmed workloads and one partition per core, partitioning the last-level cache with conventional techniques degrades throughput for 71 % of the workloads versus an unpartitioned cache (by 7% average, 25% maximum degradation), even when using 64-way caches. In contrast, Vantage improves throughput for 98% of the workloads, by 8% on average (up to 20%), using a 4-way cache.

The ZCache: Decoupling Ways and Associativity

The ever-increasing importance of main memory latency and bandwidth is pushing CMPs towards caches with higher capacity and associativity. Associativity is typically improved by increasing the number of ways. This reduces conflict misses, but increases hit latency and energy, placing a stringent trade-off on cache design. We present the zcache, a cache design that allows much higher associativity than the number of physical ways (e.g. a 64-associative cache with 4 ways). The zcache draws on previous research on skew-associative caches and cuckoo hashing. Hits, the common case, require a single lookup, incurring the latency and energy costs of a cache with a very low number of ways. On a miss, additional tag lookups happen off the critical path, yielding an arbitrarily large number of replacement candidates for the incoming block. Unlike conventional designs, the zcache provides associativity by increasing the number of replacement candidates, but not the number of cache ways. To understand the implications of this approach, we develop a general analysis framework that allows to compare associativity across different cache designs (e.g. a set-associative cache and a zcache) by representing associativity as a probability distribution. We use this framework to show that for zcaches, associativity depends only on the number of replacement candidates, and is independent of other factors (such as the number of cache ways or the workload). We also show that, for the same number of replacement candidates, the associativity of a zcache is superior than that of a set-associative cache for most workloads. Finally, we perform detailed simulations of multithreaded and multiprogrammed workloads on a large-scale CMP with zcache as the last-level cache. We show that zcaches provide higher performance and better energy efficiency than conventional caches without incurring the overheads of designs with a large number of ways.

Evaluating Bufferless Flow Control for On-chip Networks

With the emergence of on-chip networks, the power consumed by router buffers has become a primary concern. Bufferless flow control addresses this issue by removing router buffers, and handles contention by dropping or deflecting flits. This work compares virtual-channel (buffered) and deflection (packet-switched bufferless) flow control. Our evaluation includes optimizations for both schemes: buffered networks use custom SRAM-based buffers and empty buffer bypassing for energy efficiency, while bufferless networks feature a novel routing scheme that reduces average latency by 5%. Results show that unless process constraints lead to excessively costly buffers, the performance, cost and increased complexity of deflection flow control outweigh its potential gains: bufferless designs are only marginally (up to 1.5%) more energy efficient at very light loads, and buffered networks provide lower latency and higher throughput per unit power under most conditions.

SCD: A scalable coherence directory with flexible sharer set encoding

Large-scale CMPs with hundreds of cores require a directory-based protocol to maintain cache coherence. However, previously proposed coherence directories are hard to scale beyond tens of cores, requiring either excessive area or energy, complex hierarchical protocols, or inexact representations of sharer sets that increase coherence traffic and degrade performance. We present SCD, a scalable coherence directory that relies on efficient highly-associative caches (such as zcaches) to implement a single-level directory that scales to thousands of cores, tracks sharer sets exactly, and incurs negligible directory-induced invalidations. SCD scales because, unlike conventional directories, it uses a variable number of directory tags to represent sharer sets: lines with one or few sharers use a single tag, while widely shared lines use additional tags, so tags remain small as the system scales up. We show that, thanks to the efficient highly-associative array it relies on, SCD can be fully characterized using analytical models, and can be sized to guarantee a negligible number of evictions independently of the workload. We evaluate SCD using simulations of a 1024-core CMP. For the same level of coverage, we find that SCD is 13× more area-efficient than full-map sparse directories, and 2× more area-efficient and faster than hierarchical directories, while requiring a simpler protocol. Furthermore, we show that SCDs analytical models are accurate in practice.

Flexible architectural support for fine-grain scheduling

To make efficient use of CMPs with tens to hundreds of cores, it is often necessary to exploit fine-grain parallelism. However, managing tasks of a few thousand instructions is particularly challenging, as the runtime must ensure load balance without compromising locality and introducing small overheads. Software-only schedulers can implement various scheduling algorithms that match the characteristics of different applications and programming models, but suffer significant overheads as they synchronize and communicate task information over the deep cache hierarchy of a large-scale CMP. To reduce these costs, hardware-only schedulers like Carbon, which implement task queuing and scheduling in hardware, have been proposed. However, a hardware-only solution fixes the scheduling algorithm and leaves no room for other uses of the custom hardware. This paper presents a combined hardware-software approach to build fine-grain schedulers that retain the flexibility of software schedulers while being as fast and scalable as hardware ones. We propose asynchronous direct messages (ADM), a simple architectural extension that provides direct exchange of asynchronous, short messages between threads in the CMP without going through the memory hierarchy. ADM is sufficient to implement a family of novel, software-mostly schedulers that rely on low-overhead messaging to efficiently coordinate scheduling and transfer task information. These schedulers match and often exceed the performance and scalability of Carbon when using the same scheduling algorithm. When the ADM runtime tailors its scheduling algorithm to application characteristics, it outperforms Carbon by up to 70%.

An analysis of on-chip interconnection networks for large-scale chip multiprocessors

With the number of cores of chip multiprocessors (CMPs) rapidly growing as technology scales down, connecting the different components of a CMP in a scalable and efficient way becomes increasingly challenging. In this article, we explore the architectural-level implications of interconnection network design for CMPs with up to 128 fine-grain multithreaded cores. We evaluate and compare different network topologies using accurate simulation of the full chip, including the memory hierarchy and interconnect, and using a diverse set of scientific and engineering workloads. We find that the interconnect has a large impact on performance, as it is responsible for 60% to 75% of the miss latency. Latency, and not bandwidth, is the primary performance constraint, since, even with many threads per core and workloads with high miss rates, networks with enough bandwidth can be efficiently implemented for the system scales we consider. From the topologies we study, the flattened butterfly consistently outperforms the mesh and fat tree on all workloads, leading to performance advantages of up to 22%. We also show that considering interconnect and memory hierarchy together when designing large-scale CMPs is crucial, and neglecting either of the two can lead to incorrect conclusions. Finally, the effect of the interconnect on overall performance becomes more important as the number of cores increases, making interconnection choices especially critical when scaling up.

Ubik: efficient cache sharing with strict qos for latency-critical workloads

Chip-multiprocessors (CMPs) must often execute workload mixes with different performance requirements. On one hand, user-facing, latency-critical applications (e.g., web search) need low tail (i.e., worst-case) latencies, often in the millisecond range, and have inherently low utilization. On the other hand, compute-intensive batch applications (e.g., MapReduce) only need high long-term average performance. In current CMPs, latency-critical and batch applications cannot run concurrently due to interference on shared resources. Unfortunately, prior work on quality of service (QoS) in CMPs has focused on guaranteeing average performance, not tail latency. In this work, we analyze several latency-critical workloads, and show that guaranteeing average performance is insufficient to maintain low tail latency, because microarchitectural resources with state, such as caches or cores, exert inertia on instantaneous workload performance. Last-level caches impart the highest inertia, as workloads take tens of milliseconds to warm them up. When left unmanaged, or when managed with conventional QoS frameworks, shared last-level caches degrade tail latency significantly. Instead, we propose Ubik, a dynamic partitioning technique that predicts and exploits the transient behavior of latency-critical workloads to maintain their tail latency while maximizing the cache space available to batch applications. Using extensive simulations, we show that, while conventional QoS frameworks degrade tail latency by up to 2.3x, Ubik simultaneously maintains the tail latency of latency-critical workloads and significantly improves the performance of batch applications.

Dynamic Fine-Grain Scheduling of Pipeline Parallelism

Scheduling pipeline-parallel programs, defined as a graph of stages that communicate explicitly through queues, is challenging. When the application is regular and the underlying architecture can guarantee predictable execution times, several techniques exist to compute highly optimized static schedules. However, these schedules do not admit run-time load balancing, so variability introduced by the application or the underlying hardware causes load imbalance, hindering performance. On the other hand, existing schemes for dynamic fine-grain load balancing (such as task-stealing) do not work well on pipeline-parallel programs: they cannot guarantee memory footprint bounds, and do not adequately schedule complex graphs or graphs with ordered queues. We present a scheduler implementation for pipeline-parallel programs that performs fine-grain dynamic load balancing efficiently. Specifically, we implement the first real runtime for GRAMPS, a recently proposed programming model that focuses on supporting irregular pipeline and data-parallel applications (in contrast to classical stream programming models and schedulers, which require programs to be regular). Task-stealing with per-stage queues and queuing policies, coupled with a backpressure mechanism, allow us to maintain strict footprint bounds, and a buffer management scheme based on packet-stealing allows low-overhead and locality-aware dynamic allocation of queue data. We evaluate our runtime on a multi-core SMP and find that it provides low-overhead scheduling of irregular workloads while maintaining locality. We also show that the GRAMPS scheduler outperforms several other commonly used scheduling approaches. Specifically, while a typical task-stealing scheduler performs on par with GRAMPS on simple graphs, it does significantly worse on complex ones, a canonical GPGPU scheduler cannot exploit pipeline parallelism and suffers from large memory footprints, and a typical static, streaming scheduler achieves somewhat better locality, but suffers significant load imbalance on a general-purpose multi-core due to fine-grain architecture variability (e.g., cache misses and SMT).

Jigsaw: scalable software-defined caches

Shared last-level caches, widely used in chip-multi-processors (CMPs), face two fundamental limitations. First, the latency and energy of shared caches degrade as the system scales up. Second, when multiple workloads share the CMP, they suffer from interference in shared cache accesses. Unfortunately, prior research addressing one issue either ignores or worsens the other: NUCA techniques reduce access latency but are prone to hotspots and interference, and cache partitioning techniques only provide isolation but do not reduce access latency.