Cache Slice Research Articles

Imitation Learning-Based Performance-Power Trade-Off Uncore Frequency Scaling Policy for Multicore System.

As the importance of uncore components, such as shared cache slices and memory controllers, increases in processor architecture, the percentage of uncore power consumption in the overall power consumption of multicore processors rises significantly. To maximize the power efficiency of a multicore processor system, we investigate the uncore frequency scaling (UFS) policy and propose a novel imitation learning-based uncore frequency control policy. This policy performs online learning based on the DAgger algorithm and converts the annotation cost of online aggregation data into fine-tuning of the expert model. This design optimizes the online learning efficiency and improves the generality of the UFS policy on unseen loads. On the other hand, we shift our policy optimization target to Performance Per Watt (PPW), i.e., the power efficiency of the processor, to avoid saving a percentage of power while losing a larger percentage of performance. The experimental results show that our proposed policy outperforms the current advanced UFS policy in the benchmark test sequence of SPEC CPU2017. Our policy has a maximum improvement of about 10% relative to the performance-first policies. In the unseen processor load, the tuning decision made by our policy after collecting 50 aggregation data can maintain the processor stably near the optimal power efficiency state.

CACHE UTILIZATION USING A ROUTER BUFFER

Network on chip (NOC) is used to connect Multiple Cores in a tiled multicore Processor. Here each core is connected to L1 cache where LLC (L2) is shared by the entire core. The data in the Last Level Cache slices can be accessed by multiple cores simultaneously and results in congestion at the LLC and increases the access time .So in order to find a solution to this problem, Propose a congestion management Technique in the LLC that contains NOC Router with small storage (Router Buffer Cache) to keep a copy of heavily shared cache blocks. The simulation of the Multicore system can be done by Gem5 tool and evaluate its effectiveness by running a set of parallel benchmarks , Thereby Reducing the LLC access Time “

CD-Xbar: A Converge-Diverge Crossbar Network for High-Performance GPUs

Modern GPUs feature an increasing number of streaming multiprocessors (SMs) to boost system throughput. How to construct an efficient and scalable network-on-chip (NoC) for future high-performance GPUs is particularly critical. Although a mesh network is a widely used NoC topology in manycore CPUs for scalability and simplicity reasons, it is ill-suited to GPUs because of the many-to-few-to-many traffic pattern observed in GPU-compute workloads. Although a crossbar NoC is a natural fit, it does not scale to large SM counts while operating at high frequency. In this paper, we propose the converge-diverge crossbar (CD-Xbar) network with round-robin routing and topology-aware concurrent thread array (CTA) scheduling. CD-Xbar consists of two types of crossbars, a local crossbar and a global crossbar. A local crossbar converges input ports from the SMs into so-called converged ports; the global crossbar diverges these converged ports to the last-level cache (LLC) slices and memory controllers. CD-Xbar provides routing path diversity through the converged ports. Round-robin routing and topology-aware CTA scheduling balance network traffic among the converged ports within a local crossbar and across crossbars, respectively. Compared to a mesh with the same bisection bandwidth, CD-Xbar reduces NoC active silicon area and power consumption by 52.5 and 48.5 percent, respectively, while at the same time improving performance by 13.9 percent on average. CD-Xbar performs within 2.9 percent of an idealized fully-connected crossbar. We further demonstrate CD-Xbar's scalability, flexibility and improved performance per Watt (by 17.1 percent) over state-of-the-art GPU NoCs which are highly customized and non-scalable.

FOS: a low-power cache organization for multicores

The cache hierarchy of current multicore processors typically consists of one or two levels of private caches per core and a large shared last-level cache. This approach incurs area and energy wasting due to oversizing the private cache space, data replication through the inclusive cache levels, as well as the use of highly set-associative caches. In this paper, we claim that although this is the commonly adopted approach, it presents important design issues that can be addressed by a more energy efficient organization. This work proposes Flat On-chip Storage (FOS), a novel cache organization that, aimed at addressing energy and area on low-power processors, resolves the mentioned issues. For this purpose, FOS combines L2 and L3 cache levels into a single one, organized as a flat space, and composed of a pool of private small cache slices. These slices are initially powered off to save energy, and they are powered on and assigned to cores provided that the system performance is expected to improve. To provide fast and uniform access from the private L1 caches to the FOS’s cache slices, multiple architectural challenges are overcome, which entails the design of a custom optical network-on-chip. Experimental results show that FOS achieves significant energy savings on both static and dynamic energy over conventional cache organizations with the same storage capacity. FOS static energy savings are as much as 60% over an electrically connected shared cache; these savings grow up to 75% compared to optically connected baselines. Moreover, despite deactivating part of the cache space, FOS achieves similar performance values as those achieved by conventional approaches.

EECache

Power management for large last-level caches (LLCs) is important in chip multiprocessors (CMPs), as the leakage power of LLCs accounts for a significant fraction of the limited on-chip power budget. Since not all workloads running on CMPs need the entire cache, portions of a large, shared LLC can be disabled to save energy. In this article, we explore different design choices, from circuit-level cache organization to microarchitectural management policies, to propose a low-overhead runtime mechanism for energy reduction in the large, shared LLC. We first introduce a slice-based cache organization that can shut down parts of the shared LLC with minimal circuit overhead. Based on this slice-based organization, part of the shared LLC can be turned off according to the spatial and temporal cache access behavior captured by low-overhead sampling-based hardware. In order to eliminate the performance penalties caused by flushing data before powering off a cache slice, we propose data migration policies to prevent the loss of useful data in the LLC. Results show that our energy-efficient cache design (EECache) provides 14.1% energy savings at only 1.2% performance degradation and consumes negligible hardware overhead compared to prior work.

Accelerator Memory Reuse in the Dark Silicon Era

Accelerators integrated on-die with General-Purpose CPUs (GP-CPUs) can yield significant performance and power improvements. Their extensive use, however, is ultimately limited by their area overhead; due to their high degree of specialization, the opportunity cost of investing die real estate on accelerators can become prohibitive, especially for general-purpose architectures. In this paper we present a novel technique aimed at mitigating this opportunity cost by allowing GP-CPU cores to reuse accelerator memory as a non-uniform cache architecture (NUCA) substrate. On a system with a last level-2 cache of 128kB, our technique achieves on average a 25% performance improvement when reusing four 512 kB accelerator memory blocks to form a level-3 cache. Making these blocks reusable as NUCA slices incurs on average in a 1.89% area overhead with respect to equally-sized ad hoc cache slices.

Towards hybrid last level caches for chip-multiprocessors

As CMP platforms are widely adopted, more and more cores are integrated on to the die. To reduce the off-chip memory access, the last level cache is usually organized as a distributed shared cache. In order to avoid hot-spots, cache lines are interleaved across the distributed shared cache slices using a hash function. However, as we increase the number of cores and cache slices in the platform, this also implies that most of data references go to remote cache slices, thereby increasing the access latency significantly. In this paper, we propose a hybrid last level cache, which has some amount of private space and some amount of shared space on each cache slice. For workloads with no sharing, the goal is to provide more hits into the local slice while still keeping the overall miss rate low. For workloads with sufficient sharing, the goal is to allow more sharing in the last-level cache slice. We present hybrid last-level cache design options and study its hit/miss rate behavior for a number of important server applications and multi-programmed workloads. Our simulation results on running multi-programmed workloads based on SPEC CINT2000 as well as multithreaded workloads based on commercial server benchmarks (TPCC, SPECjbb, SAP and TPCE) show that this architecture is advantageous especially since it can improve the local hit rate significantly while keeping the overall miss rate similar to the shared cache.

Victim Replication

In this paper, we consider tiled chip multiprocessors (CMP) where each tile contains a slice of the total on-chip L2 cache storage and tiles are connected by an on-chip network. The L2 slices can be managed using two basic schemes: 1) each slice is treated as a private L2 cache for the tile 2) all slices are treated as a single large L2 cache shared by all tiles. Private L2 caches provide the lowest hit latency but reduce the total effective cache capacity, as each tile creates local copies of any line it touches. A shared L2 cache increases the effective cache capacity for shared data, but incurs long hit latencies when L2 data is on a remote tile. We present a new cache management policy, victim replication, which combines the advantages of private and shared schemes. Victim replication is a variant of the shared scheme which attempts to keep copies of local primary cache victims within the local L2 cache slice. Hits to these replicated copies reduce the effective latency of the shared L2 cache, while retaining the benefits of a higher effective capacity for shared data. We evaluate the various schemes using full-system simulation of both single-threaded and multi-threaded benchmarks running on an 8-processor tiled CMP. We show that victim replication reduces the average memory access latency of the shared L2 cache by an average of 16%for multi-threaded benchmarks and 24%for single-threaded benchmarks, providing better overall performance than either private or shared schemes.