Hi haproxy-list! I've got 3 active haproxy servers (multiple ip dns rr + keepalived) fronting 15 backend-servers. The haproxies have a combined session rate at ~10-14k, at ~500mbps. We have a low latency application which does programmatic bidding/buying, where every ms counts, thus CPU-utilization is very important factor in keeping it down. We are setting a CPU affinity mask to the process to keep it pinned to a single core, and in haproxy 1.4.25 it uses about 40-50% CPU.
Yesterday i upgraded haproxy from 1.4.25 to 1.5.11 and noticed increased CPU-usage, which caused problems for us. Here is CPU metrics graph from yesterday, upgrade to 1.5 took place before 12, and I downgraded to 1.4 in the evening: http://imgur.com/jrVvHlw . This increase represent going from 40-50% CPU to 94-100%. Here's also some contextual weekly graph of the same metric: http://imgur.com/o4bgtM5 Finally, here's a version of our config (i've changed names/ips/passwords, but nothing else): http://paste.ubuntu.com/10548978/ I did see something in the graphs regarding fd_inuse: http://imgur.com/hTEOfMq but I'm uncertain as what to make of it. As said, the same config runs smoothly with haproxy 1.4. I have not been able to reproduce the behavior in my lab-environment, and I would very much appreciate pointers as how to troubleshoot this further, and where to look. I have a feeling that haproxy 1.5 have default values which I've not taken into consideration that might cause this. Please advice! :-) Best regards Björn Zettergren
