Aleksandar Lazic suggested [1] that the random() distribution could
be improved to avoid picking the same server twice, comparing it with
itself and choosing it, possibly even if suboptimal. His mesurements
showed that a slow server was picked more often than desired.
It turns out that we don't need extra draws to avoid this situation,
because get_server_rnd() uses chash_get_server_hash(), which supports
a server to avoid (normally used for redispatch, but usable here as
well to avoid picking the same server twice), so we can pass there the
pointer to the currently chosen best server and count on the function
to pick another one.
Before the patch, a config with 1 slow server responding in 50ms,
and N-1 fast servers responding in 5ms would give this distribution
for 100k requests at 100 concurrency:
4 srv: req rate= 11.1k/s 8 srv: req rate = 15.8k/s
srv max total srv max total
s0 63 8525 s0 29 2248
s1 27 28296 s1 17 13874
s2 28 31049 s2 17 13869
s3 29 32130 s3 18 13996
s4 18 14316
s5 17 14325
s6 16 13758
s7 16 13614
Now with the patch:
4 srv: req rate= 14.6k/s 8 srv: req rate = 16.6k/s
srv max total srv max total
s0 29 3572 s0 17 1669
s1 28 31380 s1 15 14007
s2 29 32513 s2 16 14010
s3 29 32535 s3 15 14059
s4 16 14205
s5 16 14215
s6 16 13968
s7 15 13867
It's clear that the distribution is much better and follows the ratios
of response times, with little to no difference in maximum server load,
thus being much closer to a leastconn algo. The average rate is also
better, +33% for 4 servers and +5% for 8.
Note that in order to avoid wasting CPU cycles in degenerate farm cases,
the two previous patches are strongly recommended:
OPTIM: lb-chash: don't look for more servers when there is only one
OPTIM: lb-chash: limit the scan to few consecutive nodes
[1] https://www.mail-archive.com/haproxy@formilux.org/msg47740.html
do {
prev = curr;
hash = statistical_prng();
- curr = chash_get_server_hash(px, hash, avoid);
+ curr = chash_get_server_hash(px, hash, avoid ? avoid : prev);
if (!curr)
break;