Parallel I/O Research Topics

At Clemson we are interested in exploring several topics related to the parallel I/O research using PVFS2 as a research platform. Topics involving benchmarking or data measurements are appropriate for undergraduate research, while projects involving significant design or implementation may be more appropriate for graduate level research. The following research topics have been identified as being of interest, though we are always interested in new and interesting ideas:

Redundancy Models in parallel file systems
- Mirroring with Failover (Brad Settlemyer)
- Lazy Redundancy (Parity written on file close)
- Commit Redundancy (Parity written after write completion)
Quorum infrastructure
Failure detection/Heartbeat systems
Data Distributions
- VFS distribution selection via IOCTLS
- Romio distribution selection thru hints
- Changing the distribution on an existing file
- Custom distributions for applications
- Data migration framework
File commit infrastructure
Consistency Semantics (currently implements coherence)
- VTAGS (similar to SGI load lock/store conditional)
- Storing commits atomically
- Locking subsystem in Romio
Request region overlap detection in the request scheduler
Programmable rule based request scheduler
PVFS2 Networking improvements (BMI, etc.)
- Shared memory BMI module
- WAN protocol BMI module (bulk UDP or multiple socket)
- Locking subsystem in Romio
Production level hardening of code
Server side config file options
Future work from Phil's dissertation
Performance measurements
- Distribution strip size effects
- Synthetic benchmarks/Applications (SAGE, Blast, etc.)
- Performance impacts of various redundancy schemes
- Additional improvements of collectives on servers
- Measurements of cost of locking
- Costs of Posix compliance
- Application of Phil's models to Mike's simulator
- Use of simulator to measure something and then verify it
Workload Analysis of emerging applications