跳转至

Benchmark : GUPS

Introduction

GUPS (Giga十亿 UPdates per Second) is a measurement that profiles the memory architecture of a system and is a measure of performance similar to MFLOPS.

The HPCS HPCchallenge RandomAccess benchmark is intended to exercise the GUPS capability of a system, much like the LINPACK benchmark is intended to exercise the MFLOPS capability of a computer. In each case, we would expect these benchmarks to achieve close to the "peak" capability of the memory system. The extent of the similarities between RandomAccess and LINPACK are limited to both benchmarks attempting to calculate a peak system capability.

definition of GUPS

GUPS is calculated by identifying the number of memory locations that can be randomly updated in one second, divided by 1 billion (1e9).

  • The term "randomly" means that there is little relationship between one address to be updated and the next, except that they occur in the space of one half the total system memory. (只用一半内存?)
  • An update is a read-modify-write operation on a table of 64-bit words. An address is generated, the value at that address read from memory, modified by an integer operation (add, and, or, xor) with a literal value, and that new value is written back to memory.

Extensibility

  • We are interested in knowing the GUPS performance of both entire systems and system subcomponents --- e.g., the GUPS rating of a distributed memory multiprocessor the GUPS rating of an SMP node, and the GUPS rating of a single processor.
  • While there is typically a scaling of FLOPS with processor count, a similar phenomenon may not always occur for GUPS.

Principle

Select the memory size to be the power of two such that 2^n <= 1/2 of the total memory. Each CPU operates on its own address stream, and the single table may be distributed among nodes. The distribution of memory to nodes is left to the implementer. A uniform data distribution may help balance the workload, while non-uniform data distributions may simplify the calculations that identify processor location by eliminating the requirement for integer divides. A small (less than 1%) percentage of missed updates are permitted.

Installation

Download

official web or from GitHub

Usage

需要进一步的研究学习

暂无

遇到的问题

暂无

开题缘由、总结、反思、吐槽~~

参考文献

上面回答部分来自ChatGPT-3.5,没有进行正确性的交叉校验。