EDGE: Event-Driven GPU Execution

Hetherington, Tayler; Lubeznov, Maria; Shah, Deval; Aamodt, Tor M.

doi:10.1109/pact.2019.00034

Cited by 4 publications

(3 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Parallel Computing. Graphics Processing Units (GPUs) can perform thousands of computations in parallel depending on the availability of the number of cores on the Streaming Multiprocessors (SMs) [10,12,17]. The computations are distributed on GPUs when the CPU launches an application in the form of a kernel.…”

Section: Background and Motivationmentioning

confidence: 99%

ACE-HoT: Accelerating an Extreme Amount of Symmetric Cipher Evaluations for (High-order) Avalanche Tests

Bellini,

Grados,

Rachidi

et al. 2023

Lecture Notes in Computer Science

View full text Add to dashboard Cite

In this work, we tackle the problem of estimating the security of iterated symmetric ciphers in an efficient manner, with tests that do not require a deep analysis of the internal structure of the cipher. This is particularly useful during the design phase of these ciphers, especially for quickly testing several combinations of possible parameters defining several cipher design variants.We consider a popular statistical test that allows us to determine the probability of flipping each cipher output bit, given a small variation in the input of the cipher. From these probabilities, one can compute three measurable metrics related to the well-known full diffusion, avalanche and strict avalanche criteria.This highly parallelizable testing process scales linearly with the number of samples, i.e., cipher inputs, to be evaluated and the number of design variants to be tested. But, the number of design variants might grow exponentially with respect to some parameters.The high cost of Central Processing Unit (CPU)s makes them a bad candidate for this kind of parallelization. As a main contribution, we propose a framework, ACE-HoT, to parallelize the testing process using multi-Graphics Processing Units (GPUs). Our implementation does not perform any intermediate CPU-GPU data transfers.The diffusion and avalanche criteria can be seen as an application of discrete first-order derivatives. As a secondary contribution, we generalize these criteria to their high-order version. Our generalization requires an exponentially larger number of samples, in order to compute sufficiently accurate probabilities. As a case study, we apply ACE-HoT on most of the finalists of the National Institute of Standards and Technologies (NIST) lightweight standardization process, with a special focus on the winner ASCON.

show abstract

Section: Background and Motivationmentioning

confidence: 99%

ACE-HoT: Accelerating an Extreme Amount of Symmetric Cipher Evaluations for (High-order) Avalanche Tests

Bellini,

Grados,

Rachidi

et al. 2023

Lecture Notes in Computer Science

View full text Add to dashboard Cite

show abstract

“…Finally, EDGE [20] introduces an event-driven GPU execution model that provides the basic mechanisms needed for simultaneous execution of inference and training requests, providing the first step towards a GPU capable of meeting the latency constraints of inference services while performing training. Using EDGE, a GPU is capable of quickly preempting training requests during inference load spikes.…”

Section: Related Workmentioning

confidence: 99%

Equinox: Training (for Free) on a Custom Inference Accelerator

Drumond¹,

Coulon

Pourhabibi

et al. 2021

MICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture

View full text Add to dashboard Cite

DNN inference accelerators executing online services exhibit low average loads because of service demand variability, leading to poor resource utilization. Unfortunately, reclaiming idle inference cycles is difficult as other workloads can not execute on a custom accelerator. With recent proposals for the use of fixed-point arithmetic in training, there are opportunities for training services to piggyback on inference accelerators. We make the observation that a key challenge in doing so is maintaining service-level latency constraints for inference. We show that relaxing latency constraints in an inference accelerator with ALU arrays that are batching-optimized achieves near-optimal throughput for a given area and power envelope while maintaining inference services' tail latency goals.We present Equinox, a custom inference accelerator designed to piggyback training. Equinox employs a uniform arithmetic encoding to accommodate inference and training and a priority hardware scheduler with adaptive batching that interleaves training during idle inference cycles. For a 500𝜇𝑠 inference service time constraint, Equinox achieves 6.67× higher throughput than a latency-optimal inference accelerator. Despite not being optimized for training services, Equinox achieves up to 78% of the throughput of a dedicated training accelerator that saturates the available compute resources and DRAM bandwidth. Finally, Equinox's controller logic incurs less than 1% power and area overhead, while the uniform encoding (to enable training) incurs 13% power and 4% area overhead compared to a fixed-point inference accelerator. CCS CONCEPTS• Computer systems organization → Systolic arrays; Neural networks; Cloud computing.

show abstract

“…Thermal effects on the system as a result of CPUs and GPUs manifest in different ways. A common technique employed for thermal management of such devices is effective GPU kernel management -choosing judiciously the tasks to be scheduled on GPUs and CPUs [17].…”

Section: Introductionmentioning

confidence: 99%

SLePaaS: An Embedded Platform-as-a-Service Facilitating Research on Thermal Management of Embedded Platforms

2022

View full text Add to dashboard Cite

In this paper, we present to the embedded research community an Embedded Platform as a Service facility named SLePaaS that allows researchers remote access to experimental hardware and a collection of online tools that would facilitate research on thermal management of embedded systems. SLePaaS serves a dual purpose -provides access to actual hardware for experiments on one hand and reduces experimental cost and time on the other. With this facility, researchers neither have to rely on simulation models nor purchase or customize any experimental platform for a specific problem. In addition, the platform provides tool support, specifically for researchers working in problems related to thermal management. Presently, three tools are provided by the SLePaaS platform. The first tool helps users obtain performance, power and thermal profile of applications running on different hardware platforms. The second is a thermal value predictor tool that aids task schedulers in allocating tasks to processors based on the thermal profile of applications. The third tool is a platform predictor that helps users to decide the best platform to run an OpenCL application given a target optimization objective. All the tools have been validated for their performance and accuracy against standard benchmark applications. A prototype implementation of SLePaaS along with the three tools is available for use.

show abstract

EDGE: Event-Driven GPU Execution

Cited by 4 publications

References 41 publications

ACE-HoT: Accelerating an Extreme Amount of Symmetric Cipher Evaluations for (High-order) Avalanche Tests

ACE-HoT: Accelerating an Extreme Amount of Symmetric Cipher Evaluations for (High-order) Avalanche Tests

Equinox: Training (for Free) on a Custom Inference Accelerator

SLePaaS: An Embedded Platform-as-a-Service Facilitating Research on Thermal Management of Embedded Platforms

Contact Info

Product

Resources

About