Presentation Title: Improving Cactus Parallel I/O Performance Using the F5 Data Model and Data Compression

Committee:

• Dr. Gabrielle Allen(Major Professor)
• Dr. Brygg Ullmer
• Dr. Bijaya Karki

Date: April 3, 2009
Time: 9:30 AM
Location:297, Coates Hall

Abstract:
In the past decade, CPU speed has increased much faster than disk speed. This difference means that I/O now becomes a signi?cant bottleneck on computer systems. Today’s supercomputers, like the Queen Bee machine hosted by the Louisiana Optical Network Initiative [8], typically consist of thousands of cores but have a limited storage link. The average writing speed per core of supercomputers is very slow compared to that of computers with only one or several cores. Output speed is an essential issue for parallel computing. Simulation codes modeling physical events such as black holes, hurricanes, or molecular dynamics, produce a large amount of data at run time. Slow output performance reduces the overall software performance and thus restricts the rate of scienti?c investigation. Cactus [7] is a parallel software platform for scienti?c applications, which usually produce large amount of data and operate the disk frequently. The HDF5 [3] ?le system improves disk I/O and makes data manipulating easier for Cactus users. However, the writing speed is still restricted by the factors like the storage link speed, especially on supercomputers. Sometimes the CPU has to wait until end of the operations on disk before next computation iteration, which is a waste of CPU resources. The Fiber Bundle HDF5 (F5) [3] is a data model which is based on HDF5. It has a different semantic structures of ?les and provides a number of new features for scienti?c data. It is better for post-processing such as visualization purposes. The ease of usage and practical design for scienti?c visualization attracts many users, but the I/O speed of F5 is still an issue that need tested. There are many compression applications that exist today. They are designed using different algorithms for different types of data. For example, SZIP [11] has a good compression ratio on image data. Compression helps users to save disk speed, however, it also takes more or less CPU resources for encoding and decoding. Selection of compression application need to take account of both aspects, based on the available resources and data structures. HDF5 comes with compression support. It has a compression ?lter layer that allows developers to add their own compression application. The core motivation for compression comes when we think about using the idle CPU resources to compress the data to a smaller output, which can go through the storage link faster, so users can gain both disk space and output speed. Goals. The goals of this project are to compare the disk writing speed of the two output thorns, CactusHDF5 and CactusF5, on Cactus, implement the GZIP and SZIP compression functionalities on them, and then measure the output performance gain on output speed and disk space with compression activated. This work is done as part of the NSF XiRel and Alpaca projects which are to scale simulations to a large amount of cores. I/O is an important component of those projects. In the future, we are interested in scaling up problems to run on the Blue Waters system [4], which will have over 200,000 cores.

All are invited.