To alleviate the performance degradation introduced by cryptographic operations in web-based services, dedicated hardware accelerators have emerged for calculation offloading. However, accelerators from different vendors can be endowed with variable run-time modes and optimized for different workload types. It is necessary but challenging to fully and fairly quantify the associated acceleration benefits and determine usages for cloud tasks.

To this end, we establish a common test framework oriented for cryptographic accelerators. The micro benchmarks of our framework rely on the classic OpenSSL speed tool. The macro benchmarks of our framework present a server-client architecture. The server side modifies the Nginx web server and integrates it with Redis memory storage as well as a MySQL database to mimic practical cloud workloads in a customizable manner. The client side modifies the classic ApacheBench tool to support concurrent server access as well as automatic collection and analysis of test results. Our framework supports diverse running modes and workload types to comprehensively locate and present the capability characteristics of accelerators.

Tests are conducted on three accelerators, i.e., Intel QAT, Cavium Nitrox and Exar DX2040a. These accelerators have similar performance as reported by their vendors. The result analyses and product-to-product comparisons show that our framework can benefit cloud providers, application programmers and accelerator vendors simultaneously by providing important conclusions such as whether to offload, how to offload, and potential optimizations for more efficient offloading.

为了减轻基于Web的服务中的加密操作导致的性能下降，出现了专用的硬件加速器来进行计算分流。但是，来自不同供应商的加速器可以具有可变的运行时模式，并且可以针对不同的工作负载类型进行优化。要完全公平地量化相关的加速收益并确定云任务的使用情况，这是必要的但又是挑战。

为此，我们建立了一个面向密码加速器的通用测试框架。我们框架的微基准测试依赖于经典的OpenSSL速度工具。我们框架的宏基准测试提出了服务器-客户端体系结构。服务器端修改了Nginx Web服务器，并将其与Redis内存存储以及MySQL数据库集成在一起，以可定制的方式模拟实际的云工作负载。客户端修改了经典的ApacheBench支持并发服务器访问以及自动收集和分析测试结果的工具。我们的框架支持多种运行模式和工作负载类型，以全面定位并展示加速器的功能特性。

测试在三种加速器上进行，即英特尔QAT，Cavium Nitrox和Exar DX2040a。这些加速器具有与其供应商报告的相似的性能。结果分析和产品间的比较表明，我们的框架可以通过提供重要的结论（例如是否卸载，如何卸载以及进行更有效卸载的潜在优化）而使云提供商，应用程序程序员和加速器供应商同时受益。