slide 18 of 19

Hardware configuration

• Faster processors, etc.
• Shared or separate L2 caches?
  • Separate L2 caches minimize cache contention
  • SPMD profits from shared L2 cache fetches
• More than one PCI bus?
  • More I/O devices, more I/O parallelism
  • But SMP Linux serializes OS I/O calls
• Memory interleaving?
  • Improves DRAM cache block transfers
  • Interleave unrelated to SMP parallelism