This page discusses IO/CPU/memory balance for data-intensive extreme-scale systems.
Amdahl's Laws
The Amdahl's Law, which is now over 40 years old, defines ratios in a well balanced system. A question arises: does the law still apply to today's systems, and if so, what ratios work well for data-intensive extreme-scale systems?
The law:
- Amdahl's parallelism law: If a computation has a serial part S and a parallel component P, then the maximum speedup is S/(S+P).
- Amdahl's balanced system law: A system needs a bit of IO per second per instruction per second.
- Amdahl's memory law: alpha=1: that is the MB/MIPS ratio, in a balanced system is 1.
- Amdahl's IO law: Programs do one IO per 50,000 instructions
According to some sources, see [3], [4]:
| Application type |
Typically running on |
Amdahl number (IO:CPU) |
Amdahl number (memory:MIPS) |
|---|---|---|---|
| High-performance-computing applications |
supercomputer | ~ 1E-05 |
|
| Computation-heavy data-intensive simulations |
Beowulf cluster |
~1E-03 | |
| Data-intensive analytical applications |
Beowulf cluster |
~1E-01 to ~1 |
Survey of Existing or Planned Systems
- GrayWulf: BW=0.5, MEM = 1.04
(Dell 2950 Server, 8 core, 16 GB RAM, 15x750 GB SATA disks, dual disk controllers, Infiniband) x 12. See [5]. - BlueGene: BW=0.001, MEM=0.12
References
- The 40th Anniversary of Amdahl's Law, Gene Amdahl
- Petascale Computational Systems: Balanced CyberInfrastructure in a Data-Centric World, Gordon Bell, Jim Gray, Alex Szalay
- Amdahl Numbers as a Metric for Data Intensive Computing, Alex Szalay
- Oliver Ratzesberger's blog on eBay system
- GrayWulf