[chord] dhashclient question

[Emil Sit]

On Sun, 09 April 2006 at 22:09 (-0600), Yanyan Wang wrote:
> Or another possible reason for my observed high robustness performance
> could be, my file has too small size(200 bytes) to require 7 fragments
> to recover? Thanks!

That's correct: an object that requires less than one MTU will
essentially be highly replicated.  The reason for this is that
dhash/client.C's insert method calls dhblock_chash::generate_fragment
which calculates Ida::optimal_dfrag (size, mtu) as the number
of fragments needed for reconstruction.  From utils/ida.C

  Ida::optimal_dfrag (u_long len, u_long mtu)
  {
    // Calculate best packing for fragments to minimize packet count.
    // Estimate first assuming no per-fragment overhead.
    u_long m = (len + (mtu - 1)) / mtu;
    if (m == 0) m = 1; // Handle special case if len == 0
    // Check to see that a fragment would really fit, with overhead
    u_long nomfragsize = (len / m) + (4 + m)*2;
    // If it doesn't fit (i.e. would need to chunk), might as well parallelize.
    if (nomfragsize > mtu)
      m++;
    return m;
  }

Suppose MTU \approx 1200 (accounting for up to 300 bytes of various
transport_prot, RPC header, UDP/IP header overhead); then 
m = (200 + 1199)/1200 = 1.  So any one fragment will be sufficient to
reconstruct, i.e., replication.

I think Frank's suggestion of tweaking efrags and dfrags would do
what you want.

-- 
Emil Sit / MIT CSAIL PDOS / http://pdos.csail.mit.edu/chord/