Analysis and optimization of storage IO in distributed and massive parallel high performance systems

Abstract

Although Moore’s law ensures the increase in computational power, IO performance appears to be left behind. This minimizes the benefits gained from increased computational power. Processors have to idle for a long time waiting for IO. Another factor that slows the IO communication is the increased parallelism required in today’s computations. Most modern processing units are built from multiple weak cores. Since IO has a low parallelism the weak cores will decrease system performance. Furthermore to avoid added delay of external storage, future High Performance Computing (HPC) systems will employ Active Storage Fabrics (ASF). These embed storage directly into large HPC systems. Single HPC node IO performance will therefore require optimization. This can only be achieved with a full understanding of the IO stack operations. The analysis of the IO stack under the new conditions of multi-core and massive parallelism leads to some important conclusions. The IO stack is generally built for single devices and is heavily optimized for HDD. Two main optimization approaches are taken. The first is optimizing the IO stack to accommodate parallelism. Conclusions on IO analysis shows that a design based on several parallel operating storage devices is the best approach for parallelism in the IO stack. A parallel IO device with unified storage space is introduced. The unified storage space allows for optimal function division among resources for both read and write. The design also avoids large parallel file systems overhead by using limited changes to a conventional file system. Furthermore the interface of the IO stack is not changed by the design. This is a rather important restriction to avoid application rewrite. The implementation of such a design is shown to result in an increase in performance. The second approach is Optimizing the IO stack for Solid State Drives (SSD). The optimization for the new storage technology demanded further analysis. These show that the IO stack requires revision on many levels for optimal accommodation of SSD. File system preallocation of free blocks is used as an example. Preallocation is important for data contingency on HDD. However due to fast random access of SSD preallocation represents an overhead. By careful analysis to the block allocation algorithms, preallocation is removed. As an additional optimization approach IO compression is suggested for future work. It can utilize idle cores during an IO transaction to perform on the fly IO data compression.