PETASCALE DATA STORAGE INSTITUTE (PDSI) Final Report

Gibson, Garth¹ 

[1]Carnegie Mellon University

关键词: Petascale;    data storage;    file systems;    storage protocols;    parallel network file system;    benchmark;    checkpoint;    fault tolerance;    PDSI;   

DOI  :  10.2172/1150023
RP-ID  :  DOE-CMU-ER25767
PID  :  OSTI ID: 1150023

学科分类：数学（综合）

美国|英语

来源: SciTech Connect

【 摘 要 】

Petascale computing infrastructures for scientific discovery make petascale demands on information storage capacity, performance, concurrency, reliability, availability, and manageability. The Petascale Data Storage Institute focuses on the data storage problems found in petascale scientific computing environments, with special attention to community issues such as interoperability, community buy-in, and shared tools. The Petascale Data Storage Institute is a collaboration between researchers at Carnegie Mellon University, National Energy Research Scientific Computing Center, Pacific Northwest National Laboratory, Oak Ridge National Laboratory, Sandia National Laboratory, Los Alamos National Laboratory, University of Michigan, and the University of California at Santa Cruz. Because the Institute focuses on low level files systems and storage systems, its role in improving SciDAC systems was one of supporting application middleware such as data management and system-level performance tuning. In retrospect, the Petascale Data Storage Institute???s most innovative and impactful contribution is the Parallel Log-structured File System (PLFS). Published in SC09, PLFS is middleware that operates in MPI-IO or embedded in FUSE for non-MPI applications. Its function is to decouple concurrently written files into a per-process log file, whose impact (the contents of the single file that the parallel application was concurrently writing) is determined on later reading, rather than during its writing. PLFS is transparent to the parallel application, offering a POSIX or MPI-IO interface, and it shows an order of magnitude speedup to the Chombo benchmark and two orders of magnitude to the FLASH benchmark. Moreover, LANL production applications see speedups of 5X to 28X, so PLFS has been put into production at LANL. Originally conceived and prototyped in a PDSI collaboration between LANL and CMU, it has grown to engage many other PDSI institutes, international partners like AWE, and has a large team at EMC supporting and enhancing it. PLFS is open sourced with a BSD license on sourceforge. Post PDSI funding comes from NNSA and industry sources. Moreover, PLFS has spin out half a dozen or more papers, partnered on research with multiple schools and vendors, and has projects to transparently 1) dis- tribute metadata over independent metadata servers, 2) exploit drastically non-POSIX Hadoop storage for HPC POSIX applications, 3) compress checkpoints on the fly, 4) batch delayed writes for write speed, 5) compress read-back indexes and parallelize their redistribution, 6) double-buffer writes in NAND Flash storage to decouple host blocking during checkpoint from disk write time in the storage system, 7) pack small files into a smaller number of bigger containers. There are two large scale open source Linux software projects that PDSI significantly incubated, though neither were initated in PDSI. These are 1) Ceph, a UCSC parallel object storage research project that has continued to be a vehicle for research, and has become a released part of Linux, and 2) Parallel NFS (pNFS) a portion of the IETF???s NFSv4.1 that brings the core data parallelism found in Lustre, PanFS, PVFS, and Ceph to the industry standard NFS, with released code in Linux 3.0, and its vendor offerings, with products from NetApp, EMC, BlueArc and RedHat. Both are fundamentally supported and advanced by vendor companies now, but were critcally transferred from research demonstration to viable product with funding from PDSI, in part. At this point Lustre remains the primary path to scalable IO in Exascale systems, but both Ceph and pNFS are viable alternatives with different fundamental advantages. Finally, research community building was a big success for PDSI. Through the HECFSIO workshops and HECURA project with NSF PDSI stimulated and helped to steer leveraged funding of over $25M. Through the Petascale (now Parallel) Data Storage Workshop series, www.pdsw.org, colocated with SCxy each year, PDSI created and incubated five offerings of this high-attendance workshop. The workshop has gone on without PDSI support with two more highly successfully workshops, rewriting its organizational structure to be community managed. More than 70 peer reviewed papers have been presented at PDSW workshops.