Scaling Datacenter Accelerators with Computation Reuse Architectures

Authors: Adi Fuchs, David Wentzalff. Princeton University
Venue: ISCA 2018

Being the third paper at ISCA-18 that exploits input redundancy in one way or other (after EVA2 and Euphrates), COREx (COmputation-REuse Accelerators) proposes an effective idea to improve speedup and energy efficiency of datacenters. The paper is motivated by the manifestation of Zipf's law in data center workloads such as internet traffic and data compression.  As the paper title suggests, COREx stores the outputs and inputs of common kernels, and skips computation by sending the stored output to the host, if the current input is the same as stored input. They define the storing step as "memorization" Trading communication for computation, this work is the exact opposite of AMNESIAC (published at ASPLOS-17), which trades computation for communication.

They define 3 constraints that needs to be satisfied in-order for memorization to be successful. (1) Correct results: Memorization must produce the same result as a non-memorized system. For example, a search query like "where am I" depends not just on the input, but also depends on the location of the user, (2) Constrained storage: understanding the variance of inputs is important in designing the storage system required for memorization, (3) Cost-worthy reuse: for memorization to be effective, the cost of storing, looking-up and fetching should be less than doing the actual computation.

The COREx architecture consists of 3 main components: 1. Input Hashing Unit (IHU), 2. Input Look-up Unit (ILU), and 3. Computation History Table (CHT). IHU constructs two hashes from input blocks, index hash and pointer hash. ILU acts as a cache for CHT. ILU indexed based on index hash, and the pointer hash is stored in the data array, which acts as index to the CHT. CHT stores the <input, output> tuple. The input part is compared, and if it's a hit, the output is sent to the host, thereby avoiding computation and hence saving energy and time.


Comments

Post a Comment

Popular posts from this blog

Fundamental Latency Trade-offs in Architecting DRAM Caches (Alloy Cache)

Authors: Moinuddin K. Qureshi and Gabriel H. Loh Venue:    MICRO 2012 The authors present a new architecture for DRAM cache which targets optimizing latency rather than hit rate. The authors show that for every optimization made to improve hit rate, one must also consider the added cache latency, thus resulting in a new "break even hit rate". The authors compare to three key baselines: an infeasibly large SRAM-tag array, the prior "LH-Cache", and a ideal latency-optimized cache. They find that "unoptimizations" improve latency, while minimally impacting hit-rate. Their final result is a cache which boosts performance 35% improvement over the baseline, and nearly 20% better than the previous LH-cache. The specific contributions are: Moving to a direct mapped cache provides significant improvement (by reducing latency) Combining tag-store and data-store into a single entity improves performance 21% Implementing a small per-core predictor enhances per...
Read more

ZCOMP: Reducing DNN Cross-Layer Memory Footprint Using Vector Extensions

Authors: Berkin Akin, Zeshan A. Christi, and Alaa R. Alameldeen Venue: MICRO, 2019 With accelerators dominating the deep learning space in architecture conferences, this paper stands out as it focuses on reducing DNN inference/training overhead while using a CPU. An obvious question is, with the wide-spread use of GPUs/TPUs for deep learning, why should we focus on optimizing CPUs for deep learning. In a recent paper published at HPCA 2019, Facebook claims that CPUs are preferred for applications where a tight integration is required between DNN and non-DNN tasks. Also, Intel's recent AVX512 has specialized support for DNNs in the form of new instructions called Vector Neural Network Instructions (VNNI). Broadly, optimizing DNNs can be viewed from two different perspectives; computation, and communication. This paper targets reducing communication overhead, more specifically, activation or feature-map communication overhead by compressing them. Compressing activations/weights...
Read more

AutoFDO: Automatic Feedback-Directed Optimization for Warehouse-Scale Applications

Authors: Dehao Chen, David Xinliang Li, Tipp Moseley Venue:    CGO 2016 This paper presents AutoFDO, a system used which profiles warehouse-scale applications, and applies feedback to the compilation for the next release. AutoFDO works by profiling and storing profiles in an aggregate database, annotating the profiles via an intermediate representation, and finally providing feedback. On average, this technique boosts performance by around 10% and works well even with stale releases. For profiles to be useful, they must be in an intermediate representation. This is built from binary-level profile, which uses LBR to map instruction frequencies. By using these frequencies, program counters, and the source, a source profile can be generated. The source profile can be used to build an annotated call-graph with edge frequencies, with some inaccuracy. These annotated call-graphs can then be used as feedback to the compiler. Because of the size of the applications and scale of d...
Read more