Quote from: disq on June 16, 2011, 10:01:25 AM

Quote from: kodess on June 16, 2011, 08:39:47 AM

Quote from: Xer on June 16, 2011, 07:29:02 AM

Quote from: swusc2 on June 16, 2011, 06:50:01 AM

**********PSA**********

This is in your best interested to distribute yourselves out.

Whats stopping btcguild from using round robin or some other distributing algoritm when going over? that would solve the single server overload issue directly.
Deepbit points you to the closest server when you use their adress so it is possible to do.

Edit. everyone mining with the guiminer or other mining softwartes with hardcoded addresses would have to change miners to use those links.

he tried round robin, it overloaded the balancer...

well, the round-robin doesn't need to be smart, he could use vanilla DNS for it without any balancer. can't control it automatically and fast though (well, unless very short TTLs are used along with some bash-fu to check availability/update the main dns record) but at least wouldn't lose mining time and the eventual hashing power drop could be avoided. (of course, port 80 needs to be redirected to the main webserver which needs to run in a subdomain, maybe 'www')

Round robin at the DNS level does not work. Miner clients are dumb and re-query the DNS every connection, rather than resolve the host at the start. This causes them to bounce between servers randomly when the DNS round robin was enabled, causing work to be sent everywhere and get rejected. To date, we've tried:
DNS Round Robin - Tossed a miner between servers randomly, causing rejects.
HAProxy Load Balancing - Can't handle that many connections being proxied to remote servers.
Custom DNS Server program - DNS queries are peer-to-peer, so "sticking" a client ended up sticking a chain of DNS servers. OpenDNS and Google DNS are distributed, so a miner was sometimes being shot to different servers based on where their DNS query was routed at the time.

The final solution, which is what most people would use is to consolidate the servers to a single data center, put a HAProxy/Pen/Pound server in front of the nodes, and have it do local load balancing. However, this means that if you have connection problems to one server, you have them with every server.