Automatic and transparent offloading of computation to near-data processing units

在现代计算机系统中，在存储器层次结构(包括主存储器和存储器)或输入/输出设备和处理器之间移动要处理的数据浪费了大量的能量和时间。近数据处理(NDP)体系结构已经出现，其中有专门的处理单元位于存储数据的位置附近，例如智能SSD、智能存储器，或数据接收/发送的位置，例如在网络接口控制器上-SmartNIC。与从系统中的中央处理器请求数据相比，NDP可以更快地访问数据，并且能源成本低得多，但使现有应用程序适应这些体系结构需要付出相当大的努力，并且经常必须针对每种新的系统配置进行重新工作。我的研究项目的目标是找到一种高效的解决方案，以自动转换程序，使其在近数据处理系统上运行，而程序员只需要最少甚至不需要付出任何努力，因此是透明的。它扩展了现有的技术，如即时编译、数据流分析和性能分析，以计算以下问题的答案：在给定的NDP处理单元上运行程序的哪些部分将受益？如何以最小的时间开销将程序的执行移入和移出NDP加速器？这些方法将在现有的商用NDP硬件上进行测试，如系统内存和带有内置处理器的固态磁盘。该过程将涉及对手动移植到此硬件的软件进行基准测试，以确定哪些部件从卸载中受益最大，确定这些部件的特征并实施自动检测和卸载，然后在更广泛的软件上进行测试，以测试所发现的方法的好处。

A lot of energy and time is wasted in modern computer systems on moving data to be processed between the memory hierarchy (which include main memory and storage) or input/output devices and the processor. Near-data processing (NDP) architectures have emerged, where there are specialized processing units located near to where the data is stored, e.g., smart SSD, smart memory, or where the data is received/sent, e.g. on a Network Interface Controller - smartNIC. NDP can access the data quicker and with a lot lower energy cost than requesting it from the central processor in the system, but adapting existing applications to these architectures requires considerable effort and often has to be reworked for each new system configuration.The goal of my research project is to find an efficient solution for automatically transforming programs to run on near-data processing systems with minimal to no effort required from the programmer, hence transparently. It expands upon established techniques such as just-in-time compilation, data flow analysis and performance analysis to compute answers to questions: Which parts of a program would benefit from running on a given NDP processing units?, How to move the execution of the program with minimum time overhead to and from the NDP accelerator?The approaches will be tested on existing commercially-available NDP hardware, like system memory and solid state disks with built-in processors. The process will involve benchmarking software ported manually to this hardware to determine which parts benefit most from offloading, characterizing those parts and implementing automatic detection and offloading of them, and then testing on a wider array of software to test the benefits of the discovered approaches.