I work in remote sensing. Typical processing is a pipeline where raw data from a satellite is processed in chunks to get to some useful product. This falls into the embarrasingly parallel category, but new satellites have >2x more bits in each channel, >2x more channels, and >2x more pixels, so the size of the chunks is >8x bigger, which means you are looking at bigger storage, faster pipes, and more clock time to get processing done. 10 years ago we worked a 512x512 images was about all we could handle. Now we are looking at 4kx8k images, so a couple orders of magnitude larger. Our processors have gone from 30 to 3000 mhz, but processing time per pixel has only decreased by a factor of 10 as I/O becomes the bottleneck. People are looking at ways to transmit data in compressed form, but then you need to find ways to quickly get at a specific small region (e.g., to look at a time series of observations where there is ground truth). The other problem is that many the real-world clusters are lucky to get 50% uptime. The one down the hall was fried when the A/C died. They fixed all that, took a couple weeks to get a new A/C installed, and then a cable to the RAID stopped working, so now they have to get the cable and hope the files weren't damaged. You hear the success stories from people who have been lucky with A/C hardware, etc., but there are also lots of cluster owners who are swamped by the upkeep and or poorly maintained physical plant (power problems, A/C, etc.). -- George White <aa056 at chebucto.ns.ca> <gnw3 at acm.org> 189 Parklea Dr., Head of St. Margarets Bay, Nova Scotia B3Z 2G6