It also seems wasteful to keep re-hashing the host:port portion of the key.  The hash itself should also be inlined because it's called so many times.

Roughly what I have in mind (using crc32):

unsigned int crc = ~0;
char *p;

/* use host:port for base hash value */
spprintf(&key, 0, "%s:%d", mmc->host, mmc->port);
for (p = mmc->host; *p; p++)
    CRC32(crc, *p);
efree(key);

/* hash each successive point by combining the previous hash with the round number */
for (i = 0; i < points; i++)
{
    /* hash the 4 least significant bytes of i */
    CRC32(crc, i >> 24);
    CRC32(crc, i >> 16);
    CRC32(crc, i >> 8);
    CRC32(crc, i);

    state->points[state->num_points + i].server = mmc;
    state->points[state->num_points + i] = ~crc;
}

Typo in previous comment:

for (p = mmc->host; *p; p++)

Should be:

for (p = key; *p; p++)

This bug has been fixed in CVS.

In case this was a documentation problem, the fix will show up at the
end of next Sunday (CET) on pecl.php.net.

In case this was a pecl.php.net website problem, the change will show
up on the website in short time.
 
Thank you for the report, and for helping us make PECL better.

I've moved the memory allocation outside the loop however changing the algorithm itself would introduce inconsistencies with people who have server farms and don't update their servers all at the same time. Other clients also use the host:port-number syntax