Efficient Memory Footprint Reduction for Java Performance

我们之前在数据压缩方面的工作显著地推进了计算机体系结构的这一分支领域，也显示了在整个系统堆栈（从缓存到内存）上应用优化数据压缩的巨大潜力。正如通常发生的那样，我们的结果也突出了一些全新的问题和机遇。首先，为了提高效率，数据压缩需要非常快，特别是解压缩部分，因为它经常在关键的执行路径上。其次，不同应用程序的可压缩性差异很大，但仍然是通用的，足以引起人们的兴趣，并且在CPU和GPGPU应用程序的60%到75%的不同应用程序中都可以观察到。第三，为了提高内存层次结构的效率，最好使用特定于领域或特定于模式的压缩算法，这些算法可能会牺牲一些压缩比，以获得可接受的压缩/解压缩开销。在这项工作中，我们的目标是将我们在数据压缩方面的经验应用到当前数据压缩应用非常有限的一个新领域——即时（JIT）编译。根据我们最初的讨论，有一个强烈的迹象表明内存占用是Java性能的一个重要关注点。在我们的项目中，我们的目标是使用数据压缩来解决这个问题。我们设想了我们研究的几个主要方向。首先，我们希望收集一组有代表性的Java应用程序，并生成它们的内存转储，其中包括JIT元数据、编译对象和堆。其次，我们将对现有的压缩算法进行广泛的分析，这些算法可能对内存消耗的主要贡献者都很有用。第三，我们将寻找潜在的特定于领域的编码，这些编码将进一步提高Java应用程序的可压缩性。第四，我们将研究在可能的情况下直接在压缩数据上运行执行的可能性。

Our prior work on data compression significantly advanced this subfield of computer architecture, and also showed the significant potential of applying optimized data compression across the whole systems stack (from caches to memory). As it commonly happens, our results also highlighted some completely new problems and opportunities. First of all, to be efficient, data compression needs to be very fast, especially decompression part as it is frequently on the critical execution path. Second, compressibility varies a lot for different applications, but still is generic enough to be interesting and is observed in 60% to 75% of different applications for both CPU and GPGPU applications. Third, to be efficient in memory hierarchy, it is better to use domain-specific or pattern-specific compression algorithms that might sacrifice some compression ratio for acceptable overhead of compression/decompression. In this work, we aim to apply our experience on data compression to a new area where the current application of data compression was very limited so far - just-in-time (JIT) compilation. Based on our initial discussion, there is a strong indication that memory footprint is an important concern for Java performance. In our project, we aim to address this problem using data compression. We envision several major directions in our research. First, we would like to collect a set of representative Java applications and generate their memory dumps that would include both the JIT metadata, compiled objects, and the heap. Second, we will perform an extensive analysis of existing compression algorithms that might be useful for each for the major contributors to the memory consumption. Third, we will look for potential domain-specific encodings that will improve compressibility of Java applications even further. Fourth, we will investigate a potential of running execution directly on compressed data when possible.