Elastic Management of Reconfigurable Accelerators

Abstract

This paper presents a runtime system for reconfigurable accelerators that supports elastic management: it enables effective sharing of accelerator resources across multiple applications. For each application, this runtime system allocates an appropriate amount of resources to satisfy its quality-of-service requirements, while minimising the overall execution time for a collection of applications. The effectiveness of this runtime system is due to a set of scheduling algorithms and strategies customised for different types of workloads. We demonstrate our approach by implementing a dynamic Monte Carlo bond options pricing design.

dfesnippets: An Open-Source Library for Dataflow Acceleration on FPGAs

“…To illustrate the productivity gains achievable by the applications components we note that recent software frameworks such as experimental compilers [12] and resource management frameworks [13] for FPGA based systems can utilise these applications directly as benchmarks. Prototypes for these projects require 4123 and 3880 lines of code respectively, while the benchmarks require 2050 and 2924 lines of code respectively.…”

“…dfesnippets includes a multi-FPGA dynamic Monte Carlo design for bond options pricing. This design is particularly useful as a case study for resource management frameworks or environments for FPGAs [13,14] as it demonstrates good scalability and performance. To accelerate the payoff evaluation for the bond option, the Monte Carlo paths and the payoff evaluation functions are implemented on the FPGA accelerator.…”

“…This makes the design easy to place and route, and therefore ideal for experimental workloads where a short iteration time is essential, for example when developing tests and benchmarks for more complex tools. The design achieves linear scalability [13] and can therefore be used to benchmark load distribution tools, scheduling strategies and cloud-like environments for heterogeneous systems, such as FPGAs.…”

“…An application component which provides a collection of full applications to be used as case-studies for the development of frameworks and tools for FPGA based programming. These applications have been used in several research projects and publications [12][13][14][15][16]. Although we will not cover this in greater detail due to lack of space, the library also contains: 1) header only C++ libraries implementing useful functionality for managing and benchmarking DFE projects ranging from timing utilities to APIs for reordering sparse matrix data in preparation for FPGA execution; 2) tools for creating and managing projects such as to compile, generate and manage multiprocess and multi-node hardware compilation, and automatically extract and tabulate resource usage and generate reports; 3) comprehensive, automated test suite, testing each design to ensure it is functionally correct and it meets timing and resource usage constraints.…”

Quality of Service Driven Runtime Resource Allocation in Reconfigurable HPC Architectures

“…Hydrogen [16] attempts to provide elasticity in multi-tenants environments for the same class of systems targeted by this work. It associates each job with a Job-Level Objective metric (e.g., execution time); each Hydrogen instance manages a pool of resources, and a resource manager allocates jobs to the resources according to a scheduling policy.…”