The Case for a Hardware Filesystem
1 online resource (77 pages) : PDF
University of North Carolina at Charlotte
As secondary storage devices get faster with flash based solid statedrives (SSDs) and emerging technologies like phase change memories (PCM), overheads in system software like operating system (OS) and filesystem become prominent and may limit the potential performance improvements. Moreover, with rapidly increasing on-chip core count, monolithic operating systems will face scalability issues on these many-core chips. Future operating systems are likely to have a distributed nature, with a separation of operating system services amongst cores. Also, general purpose processors are known to be both performance and power inefficient while executing operating system code. In the domain of High Performance Computing with FPGAs too, relying on the OS for file I/O transactions using slow embedded processors, hinders performance. Migrating the filesystem into a dedicated hardware core, has the potential of improving the performance of data-intensive applications by bypassing the OS stack to provide higher bandwdith and reduced latency while accessing disks. To test the feasibility of this idea, an FPGA-based Hardware Filesystem (HWFS) was designed with five basic operations (open, read, write, delete and seek). Furthermore, multi-disk and RAID-0 (striping) support has been implemented as an option in the filesystem. In order to reduce design complexity and facilitate easier testing of the HWFS, a RAM disk was used initially. The filesystem core has been integrated and tested with a hardware application core (BLAST) as well as a multi-node FPGA network to provide remote-disk access. Finally, a SATA IP core was developed and directly integrated with HWFS to test with SSDs. For evaluation, HWFS's performance was compared to an Ext2 filesystem, both on an FPGA-based soft processor as well as a modern AMD Opteron Linux server with sequential and random workloads. Results prove that the Hardware Filesystem and supporting infrastructure provide substantial performanceimprovement over software only systems. The system is also resource efficient consuming less than 3% of logic and 5% of the Block RAMs of a Xilinx Virtex-6 chip.
Conrad, JamesJoshi, BharatCai, Zongwu
Thesis (Ph.D.)--University of North Carolina at Charlotte, 2012.
This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s). For additional information, see http://rightsstatements.org/page/InC/1.0/.
Copyright is held by the author unless otherwise indicated.