SPECweb99 Result

Performance

Configuration

Notes

SUT Notes

• 1 disk for OS
• 13 disk for 'specweb' (file_set, logs),software RAID0
• - using 2MB chunk size for the raid array (/etc/raidtab: default 64k)
• set noatime,nodiratime,nocheck for specweb mount (/etc/fstab):
• - noatime/nodiratime: Do not update inode access times on this file system.
• - nocheck: No checking is done. ("no check on mount")

Operating System Notes

• ulimit -n 10000, sets number of open files, default 1024
• /proc/sys/net
• - core.optmem_max=10000000, maximum amount of option memory buffers, default 10240
• - core.rmem_default=10000000, default receive socket buffer size, default 65535
• - core.rmem_max=10000000, maximum receive socket buffer size, default 131071
• - core.wmem_default=10000000, default send socket buffer size, default 65535
• - core.wmem_max=10000000, maximum send socket buffer size, default 131071
• - core.hot_list_length=5000, maximm number of skb-heads to be cached, default 128
• - core.netdev_max_backlog=300000, number of unprocessed input packets before kernel starts dropping them, default 300
• - ipv4.tcp_max_tw_buckets=2000000, set TCP time-wait buckets pool size, default 180000
• - ipv4.tcp_rmem=30000000 30000000 30000000, maximum TCP read-buffer space allocatable, default 4096 87380 174760
• - ipv4.tcp_wmem=30000000 30000000 30000000, maximum TCP write-buffer space allocatable, default 4096 16384 131072
• - ipv4.tcp_mem=30000000 30000000 30000000, maximum TCP buffer space allocatable, default 195584 196096 196608
• - ipv4.tcp_timestamps=0, turns TCP timestamp support off, default on
• - ipv4.tcp_dsack=0, turns DSACK support off, default on
• - ipv4.tcp_sack=0, turns SACK support off, default on
• - ipv4.tcp_window_scaling=0, turns TCP window scaling support off, default on
• - ipv4.ip_forward=1, default off
• /proc/sys/vm
• - dcache_priority=1000000,lowers #entries zapped from dcache each iteration, default 0
• Kernel parameters:
• - idle=poll, idle threads poll reschedule flag, default: use HLT instruction

HTTP Software Notes

• /etc/sysconfig/tux:
• - DOCROOT=/specweb, default: /var/www/html
• - CGI_UID=99, userid for cgi scripts; default: -1
• - CGI_GID=99, groupid for cgi scripts; default: -1
• - CGIROOT=/, cgi root check; default: /var/www/html
• - TUXMODULES=CAD_u.tux, list of user-space TUX modules
• - MODULEPATH=/, path to user-space TUX modules
• /proc/sys/net
• - tux.mode_allowed=255, allowed file permission bits; default: 4
• - tux.mode_forbidden=2, forbidden file permission bits; default: 0
• - tux.logging=1, enable logging; default: off (0)
• - tux.logfile=/specweb/log/tux, TUX logfile; default: /var/log/tux
• - tux.max_connect=50000, default: 1000
• - tux.max_backlog=3000, max size of SYN backlog (TUX listening socket); default: 2048
• - tux.generate_etags=0, no ETag header; default: 1
• - tux.generate_last_mod=0, no last modified header; default: 1
• - tux.noid=1, no server ID string; default: 0
• - tux.cgi_inherit_cpu=1, CGI scripts inherit the CPU-binding; default: off (0)
• /proc/net
• - tux.0.listen.0=http://10.100.1.100:80, tie NIC to 1. TUX thread, default 0
• - tux.1.listen.0=http://10.100.2.100:80, tie NIC to 2. TUX thread, default 0
• - tux.2.listen.0=http://10.100.3.100:80, tie NIC to 3. TUX thread, default 0
• - tux.3.listen.0=http://10.100.4.100:80, tie NIC to 4. TUX thread, default 0
• /etc/tux.mime.types
• - uncomment: TUX/redirect, TUX/CGI, TUX/module

HTTP API Notes

• Archived in 'Fujitsu-Siemens-20050201-tux.tar'
• - CAD_u.c, DYN_get_script.c, dof (compiler flags)
• - config.tux, config.kernel

Network Notes

• Intel PRO/1000 MT Server Adapter
• (insmod e1000 eth0 eth1 eth2 eth3)
• - FlowControl=0 (off), default read settings from the EEPROM
• - RxDescriptors=768, default 256
• - TxDescriptors=4096, default 256
• (ifconfig eth0 eth1 eth2 eth3)
• - txqueuelen=20000, transmit queue length; default 100
• /proc/irq
• - 52.smp_affinity=00000001, lock IRQ of NIC to 1. processor; default 0xffffffff
• - 53.smp_affinity=00000002, lock IRQ of NIC to 2. processor; default 0xffffffff
• - 54.smp_affinity=00000004, lock IRQ of NIC to 3. processor; default 0xffffffff
• - 55.smp_affinity=00000008, lock IRQ of NIC to 4. processor; default 0xffffffff
• - irqbalance --noethernet; default irqbalance

Client Notes

• NDIS registry tuning parameters [HKLM\System\CurrentControlSet\Services\NDIS\Parameters\]:
• - ProcessorAffinityMask=0, default 1 (any processor)
• TCP/IP registry tuning parameters [HKLM\System\CurrentControlSet\Services\Tcpip\Parameters\]:
• - MaxHashTableSize=0x4000, sets TCB hash table size to 16384 entries, default is 512.
• - TcpTimedWaitDelay=0x3c, sets TIME_WAIT parameter to 60 seconds (non-RFC 1122), default is 240.
• - MaxFreeTwTcbs=400000, sets maximum number of TIME_WAIT tcbs to 400,000.
• - MaxFreeTcbs=0x16000, sets number of cached (pre-allocated) TCBs, default 1000
• - MaxUserPort=0xfffe, sets maximum open ports to 65534.
• - TcpWindowSize=65535, sets TCP send/receive window size, default 8192.
• - NumTcbTablePartitions=32, sets number of independent TCB table segments, default 4.

Test Run Details

Tue Feb 15 17:15:56 2005

First published at SPEC.org on 15-Feb-2005