[CentOS] Memcache timeouts?

[Leon Fauster]

Am 22.03.2013 um 15:37 schrieb Rafał Radecki <radecki.rafal at gmail.com>:
> Hi All.
> 
> I am currently using memcache daemon in version 1.4.15 on one of my
> servers. The second one is executing php scripts which use mentioned
> memcache daemon (php53-pecl-memcache-2.2.5).
> 
> Memcache server:
> CentOS release 6.3 (Final)
> 2.6.32-279.5.2.el6.centos.plus.x86_64
> uptime 184 days
> 
> PHP server:
> CentOS release 6.3 (Final)
> 2.6.32-279.19.1.el6.centos.plus.x86_64
> uptime 43 days
> 
> The memcache server gets about:
> - 250Mb/s traffic
> - 4500 connections/s
> 
> On PHP server I get sometimes timeouts, ~20-30 daily. Both servers are
> not swapping, they have free memory, cpu is ~40% used max. But I have
> found some errors in netstat -s:
> 
> Memcache server:
> TCP
> 7251 failed connection attempts (incrementing)
> 56447 connection resets received (incrementing)
> TcpExt
> 36 packets pruned from receive queue because of socket buffer overrun
> (not incrementing)
> 2820 packets collapsed in receive queue due to low socket buffer (not
> incrementing)
> 459479 connections reset due to unexpected data (incrementing)
> 
> PHP server:
> TCP
> 540 failed connection attempts (incrementing)
> 381066 connection resets received (incrementing)
> TcpExt
> 7194 packets pruned from receive queue because of socket buffer
> overrun (incrementing)
> 250104 packets collapsed in receive queue due to low socket buffer
> (incrementing)
> 2447931 connections reset due to unexpected data (incrementing)
> 
> I have tuned tcp/ip a bit:
> 
> net.ipv4.tcp_fin_timeout = 5
> net.ipv4.tcp_tw_reuse = 1
> net.core.somaxconn = 1024
> net.ipv4.tcp_max_syn_backlog = 4196
> net.core.netdev_max_backlog = 4196
> net.ipv4.tcp_sack = 0
> net.ipv4.ip_local_port_range = 8192 65534
> 
> net.core.rmem_max = 8388608
> net.core.wmem_max = 8388608
> net.ipv4.tcp_rmem = 4096 65536 8388608
> net.ipv4.tcp_wmem = 4096 65536 8388608
> 
> and set txqueuelen for interfaces (1Gb/s) to 5000.
> 
> but where to look next? May it be so that the php client version is
> buggy? Any other tips?



just supposing - are you using httpd? -> listenbacklog 

--
LF