Back-of-the Envolope Calculations

Owned by John Ousterhout
Last updated: Dec 07, 2009Version comment

This page contains a collection of rough estimates of various interesting properties of a RAMCloud system:

  • Rate of crashes: every 15 minutes (assumes 10,000 servers, 3 crashes/server/year)
  • Time to completely recover a crashed server (assumes 64 GB of memory must be transferred to a new server over the network):
    • 10 minutes @ 1Gb/s
    • 1 minute @ 10Gb/s
  • Time to recover hash table for a crashed server (64 GB memory, 8 bytes/object):
    • 100-byte objects => 640M objects => 5GB => 5 secs. @ 10Gb/s, 50 secs @ 1Gb/s
    • 1000-byte objects => 64M objects => 0.5GB => .5 sec @ 10Gb/s, 5 secs @ 1 Gb/s
  • Number of segments (assuming 8MB/segment):
    • 8000 segments per server (assumes 64GB memory/server)
    • 80M segments in a RAMCloud (assumes 10,000 servers)