Critical issues remain to utilize the ASCI platforms on SBSS workloads
- Application/architecture mapping ? performance
- Most efficient applications (uniform grid sPPM) get 30% of peak
- Typical block structured applications performance ~10% of peak
- Typical arbitrarily connected grid applications ~1-3% of peak
- The processor speed improvements are much faster than memory speed improvements
- The processor speed improvements are an astounding 60%/yr
- Memory improvements: density (60%/yr), bandwidth (~10%/yr) and latency 6%/yr
- Modern software techniques (object oriented languages) and modern data abstractions (arbitrarily connected grids) together present modern architectures pathologically poor usage patterns
- Applications scaling issues poorly understood
- What future MPI constructs are necessary?
- Would low latency GET/PUT mechanisms be better?
- I/O, Networking, Visualization and archival scaling critical issues
- Poor platform utilization due to lack of batch scheduling methodologies at scale
- Limited policy options available to decision makers make for hard choices