static void sortHostArray(HostId *blockHosts,
Int32 numBlocks,
Int32 replication,
const NAString &randomizer)
{
// the hdfsGetHosts() call randomizes the hosts for 1st, 2nd and 3rd replica etc.
// for each call, probably to get more even access patterns. This makes it hard
// to debug the placement algorithm, since almost no 2 query plans are alike.
// Replace the random method of hdfsGetHosts with a pseudo-random one,
// based on the file name. With no randomization we would put a bigger load
// on hosts with a lower id.
// we have replication * numBlocks entries in blockHosts, with entry
// (r * numBlocks + b) being the rth replica of block #b.
if (replication > 1 && replication <= 10)
{
UInt32 rshift = (UInt32) randomizer.hash();
for (Int32 b=0; b<numBlocks; b++)
{
// a sorted array of HostIds for a given block
HostId s[10];
// insert the first v
s[0]=blockHosts[b];
for (Int32 r=1; r<replication; r++)
{
HostId newVal = blockHosts[r*numBlocks + b];
// replication is a small number, bubblesort of s will do...
for (Int32 x=0; x<r; x++)
if (newVal < s[x])
{
// shift the larger values up by 1
for (Int32 y=r; y>x; y--)
s[y] = s[y-1];
// then insert the new value
s[x] = newVal;
break;
}
else if (x == r-1)
// new value is the largest, insert at end
s[r] = newVal;
} // for each replica host of a block
// now move sorted values in s back to blockHosts,
// but shift them by rshift mod replication
for (Int32 m=0; m<replication; m++)
blockHosts[m*numBlocks + b] = s[((UInt32) m + rshift + (UInt32) b) % replication];
} // for each block b
} // replication between 2 and 10
} // sortHostArray
