Garth Gibson – Talk

Complexity Issues in Data Intensive High End Computing

Panel Chair: Dr. Garth Gibson, Assoc Prof, CMU, and CTO, Panasas, USA

Take it as a given that computational speeds are going up and that this will lead to more data being created, accessed and managed. So what is the big deal? Clusters get bigger, applications get bigger, so why would storage getting bigger be any harder? Could it be the 9s? Having every byte of tera- and increasingly petabyte stores available to all nodes with good performance for all but minutes a year, when files and volumes are parallel apps on the storage servers, might be a higher standard than compute nodes are held to.
Or perhaps it is deeper and deeper writebehind and readahead, and more and more concurrency, needed to achieve the ever larger contiguous blocks that are needed to minimize seeks in ever wider storage striping. Or maybe Amdahl’s law is hitting us with the need to parallelize more and more of the metadata work which has been serial and synchronous for correctness and error code simplicity in the past. Or maybe parallel file systems developers have inadequate development tools in comparison to parallel app writers. Or perhaps storage system developers are just wimps.

This panel will try to expose some heat and light on these concerns.

Questions for panel members:

  1. Bandwidth: In the next decade is the bandwidth transferred into or out of one "high end computing file system" (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", as a result of the expected increase in computational speed in its client clusters/MPPs, and why?
  2. Spindle Count: In the next decade is the number of magnetic disks in one "high end computing file system" (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", as a result of the expected increase in computational speed in its client clusters/MPPs, and why?
  3. Concurrency: In the next decade is the number of concurrent streams of requests applied to one "high end computing file system" (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", as a result of the expected increase in concurrency in client clusters/MPPs, and why?
  4. Seek Efficiency: In the next decade is the number of bytes moved per magnetic disk seek in one "high end computing file system" (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", as a result of the expected increase in computational speed in its client clusters/MPPs, and why?
  5. Failures: In the next decade is the number of independent failure domains in one "high end computing file system" (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", and why?
  6. Coping with Complexity: If you have answered (c) one or more times, please explain why these large increases are not going to increase the complexity of storage software significantly? Are you relying on the development of any currently insufficient technologies, and if so, which?
  7. Development Time Trends: If complexity is increasing in high end computing file systems, is the time and effort required to achieve acceptable 9s of availability at speed (a) going down 10X or more, (b) staying about the same, (c) going up 10X or more, or (d) "your answer here", and why? Are you relying on the development of any currently insufficient technologies, and if so, which?