Tuning Linux Windows and Firebird for Heavy Workload

TUNING LINUX, WINDOWS AND
FIREBIRD FOR HEAVY WORKLOAD
Alex Kovyazin,
IBSurgeon
Firebird Tour 2017: Performance Optimization
Prague, Bad Sassendorf, Moscow

Firebird 2017 Tour: Performance Optimization
• Firebird Tour 2017 is organized by Firebird Project,
IBSurgeon and IBPhoenix, and devoted to Firebird
Performance
• The Platinum sponsor is Moscow Exchange
• Tour's locations and dates:
• October 3, 2017 – Prague, Czech Republic
• October 5, 2017 – Bad Sassendorf, Germany
• November 3, 2017 – Moscow, Russia

• Platinum Sponsor
• Sponsor of
• «Firebird 2.5 SQL Language Reference»
• «Firebird 3.0 SQL Language Reference»
• «Firebird 3.0 Developer Guide»
• «Firebird 3.0 Operations Guide»
• Sponsor of Firebird 2017 Tour seminars
• www.moex.com

• Replication, Recovery and
Optimization for Firebird
and InterBase since 2002
• Platinum Sponsor of
Firebird Foundation
• Based in Moscow, Russia
www.ib-aid.com

Agenda
• Real customers with big databases
• Hardware they use
• OS tuning
• CPU
• RAM
• IO
• Network
• Firebird configuration

Customer 1: http://klinikabudzdorov.ru
• BudZdorov
• Medical centers and
hospitals in Moscow,
Saint-Petersburg and
major cities in Russia
• 17 departments
• 365 days per year, from 8-
00 to 21-00

ERP with Firebird in BudZdrorov
Central
Database
Replica of
Central
Database
Department’s serverCentral server
Standby for central server
Department’s
DB
Replica dept server
Department’s
DB
17departments

BudZdorov: Central database
• Size = 453 Gb
• Daily users = from 700 to 1800 (peak)
• Hardware server
• OS = Linux CentOS 6.7
• Firebird 2.5 Classic + HQbird
• Client-server, connected through optic with departments
• With async replica on the separate server

Tuning Linux Windows and Firebird for Heavy Workload

Customer 2: Customer revoked permission to publish
information 
• Customer #2
• Repair services for
xxxxx across Russia
• 365 days per year,
24x7, with 1 hour
maintenance every
day


Customer #2: Central Database
• Size = 250Gb
• Daily users from 500 to 1000 (peak)
• Hardware server
• Windows 2012R2
• Firebird 3
• Middleware (web)

Performance problems – as usual
• Long running active transactions
• Garbage collection is blocked for hours and even days
• Badly written SQLs in applications
• Peaks of load
• People are mostly sick during the winter
• Railroad has peak of loads
• Anti-failure approach
• Replica with 1 minute delay

Tuning goals
1. Tune for throughput first, then, if possible, for response
time
1. During the day users are Ok with performance
2. Problems occur only during periods of high load
2. Tune OS to get appropriate results from the powerful
hardware

General requirements for high load server
1. Not a Primary/Backup Controller/Small Business Server (Windows)
2. No Exchange (store.exe and MSSQL inside) or Sharepoint (MSSQL
inside) or dedicated MSSQL
• Each MSSQL should be restricted in memory usage
3. Not a File Server/Print Server/Terminal Server/Web server
4. If it is virtual machine, it should be really fast
5. If there is your middleware - does it benefit from being on the same
server (i.e., local protocol)?
1. If not, put it on another server
2. If yes, make sure to allocate resources
Dedicated server means dedicated!

Hardware configuration in BudZdorov
• Server model: HP ProLiant DL380p Gen8 2x Xeon(R) CPU E5v2 @
2.60GHz
• 2 processors* 6 physical cores * 2 HyperThreading = 24 cores
• RAM 384Gb
• Disks:
• RAID10 array on SSDs – 680Gb – for database
• Tmpfs on SSD -158Gb
• SAN on SAS15k - 1.8Tb
• External mounted backup partition for 1.4Tb
• Network
• BroadCom NetXtreme BCM5719 Gigabit Ethernet PCIe

Hardware configuration in Customer#2
• Server model: Dell PowerEdge R810, 2x Xeon(R) CPU E5-2630
v4
• 24 cores
• RAM 256Gb
• Disks:
• RAID1 array on SSDs – 480Gb – for database
• OS on SAS15K - 160Gb
• Network
• Broadcom 57810, 10Gb/sec

CPU
• How to improve CPU utilization?
• How can we improve distribution of load between cores?

CPU at Linux
• irqbalance
• yum install -y irqbalance && chkconfig irqbalance on &&
service irqbalance start
• Result: better CPU load distribution, increased throughput

CPU at Windows
• Windows: only CPU_AFFINITY in Firebird configuration
• Result: some cores can be excluded from Firebird usage
(reserved for middleware/other services), less conflicts,
slightly better throughput

RAM Tuning
• How to effectively use available RAM?
• How to avoid swapping?
• Firebird settings:
• DefaultDBCachePages – page buffers cache
• FileCacheSystemThreshold – limit to use/not use file cache
• TempCacheLimit – memory space for sorting

Paged Memory
File Cache/File Buffers and Page Cache
Tuning RAM: 3 types of memory
Database fileDatabase file
Page Buffers
Kernel

Paged Memory
File Cache
RAM in case of Big Databases and Big Caches
Database fileDatabase file
Page Buffers
Kernel
Competition

OS Memory Manager vs Firebird
• If Page Buffers is more than Paged Memory, OS Memory
Manager tries to send it to swap
• Race for resources between Paged Memory and File
Cache leads to swapping

Tuning RAM on Linux
• On Linux RedHat/CentOS file cache is not limited by
default
vm.pagecache = 100 #default
• For Classic – it is more or less fine, since it uses file cache
heavily
• For SuperServer it is not great, since SS 3.0 can use many
page buffers
Recommendation is to limit file cache to 40-50%:
vm.pagecache = 50

Tuning RAM on Linux
• We know that database should be kept in RAM: need to reduce
swapping!
• vm.swappiness = 10
• vm.dirty_ratio = 60
• vm.dirty_background_ratio = 2
• vm.min_free_kbytes = 1048576

Tuning RAM at Windows
•Windows Memory Manager has the following
default scenario of using RAM:
50% paged memory
41% file cache
9% kernel
•Memory distribution can be changed in registry/role
settings
Tip: use RAMMap tool to
see memory allocation

Recommendations for RAM on Windows
• Page Buffers must be < Paged Memory (50% of RAM by
default)
• %% can be changed on Windows level
• File Cache should be On
• For Classic and SuperClassic without exceptions
• For SuperServer with databases with size more than RAM > 2x
• File Cache should be enough to keep frequently requested
parts of database
• Firebird by default has file cache enabled: condition is
DefaultDBCachePages < FileSystemCacheThreshold

When can we disable File Cache?
• File Cache can be
disabled for SuperServer
for
• Read Only databases
• For database which fits
into Page Buffers with very
low % of writes
• For databases on SSD
with small % of writes
• Test it!

Paging file tuning
• In case of balanced settings for Page Buffers and enabled
File Cache, and in case of RAM > 32Gb, page file can be
limited to 16Gb.
• Page file will work fast on SSD – but not on the SSD with
database!
• Monitor life span of SSD!

Linux: general recommendations
• Centos
• Linux version 2.6.32-642.13.1.el6.x86_64 (mockbuild@c1bm.rdu2.centos.org)
(gcc version 4.4.7 20120313 (Red Hat 4.4.7-17) (GCC) ) #1 SMP Wed Jan 11
20:56:24 UTC 2017 – not so good, better choose newer OS version
• Use fresh and popular Linux distributions: Ubuntu 16+ Server and
CentOS 7+
• Use server version of Linux distributions – it has already tuned
limits for number of open files

Linux: file and process limits
# increase max user processes ulimit
(-u) 1291632
# Increase size of file handles and inode
cache
fs.file-max = 2097152

Process forking is set to unlimited
• [root@mskv-cbd-new limits.d]# cat
/etc/security/limits.d/90-nproc.conf
• * soft nproc unlimited
• root soft nproc unlimited
• [root@mskv-cbd-new security]# sed -e 's/^[ t]*//'
/etc/security/limits.conf | grep "^[^#;]" | sort
• firebird - nofile 32768
• * soft core unlimited

/etc/xinetd.conf – the most important
# cps = 25 30 ==> configures xinetd to allow
#no more than 25 connections PER SECOND to any given
service. If this limit is reached, the service is
retired for 30 seconds.
cps = 1500 10
# Sets the maximum number of requests xinetd can
handle at once.
instances = UNLIMITED
# per_source — Defines the maximum number of
#instances for a service per source IP address
per_source = UNLIMITED

IO
•For RAID
•Write-Back
•Enable cache
•Setup ratio Reads/Writes according your load
•BBU!
• SSD!

IO on Linux: File System and Barriers
• Ext4
Since we have RAID and Enterprise SSDs with power loss
protection(and high quality hardware):
Barrier = 0 (disabled)

Disk IO on BudZdorov
• SSDs deliver high speed: 242Mb/sec

IO on Windows
• Enable disk cache (it does not work on Primary Disk
Controller)

Temp space on RAM/SSD?
• TempCacheLimit – by default it is very low, increase it!
• Temp files are created in %TEMP% or /tmp or in
TempDirectories
• Big TempCacheLimit allows to avoid temp files
• However, we still need big TempDirectories to
create/restore indices

Network
# Increase number of incoming connections
net.core.somaxconn = 4096
# Increase number of incoming connections backlog
net.core.netdev_max_backlog = 65536
# Increase the maximum amount of option memory buffers
net.core.optmem_max = 25165824
# Increase the tcp-time-wait buckets pool to prevent simple DOS attacks
net.ipv4.tcp_max_tw_buckets = 1440000
net.ipv4.tcp_tw_recycle = 1
net.ipv4.tcp_tw_reuse = 1

Network
#Number of times SYNACKs for passive TCP connection.
net.ipv4.tcp_synack_retries = 2
#Allowed local port range
net.ipv4.ip_local_port_range = 2000 65535
#Protect Against TCP Time-Wait
net.ipv4.tcp_rfc1337 = 1
#Decrease the time default value for tcp_fin_timeout connection
net.ipv4.tcp_fin_timeout = 15
#Decrease the time default value for connections to keep alive
net.ipv4.tcp_keepalive_time = 300
net.ipv4.tcp_keepalive_probes = 5
net.ipv4.tcp_keepalive_intvl = 15

Network
net.ipv4.tcp_congestion_control=htcp
net.ipv4.tcp_no_metrics_save=1
net.ipv4.tcp_moderate_rcvbuf=1
net.ipv4.tcp_slow_start_after_idle=0
net.core.rmem_default = 65536
net.core.wmem_default = 65536
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.ipv4.tcp_mem = 50576 64768 98152
net.ipv4.tcp_rmem = 4096 87380 16777216
net.ipv4.tcp_wmem = 4096 65536 16777216

Network queues
# For 24 CPU and 4 RX queues on NIC
cat > /root/scripts/rps_boot.sh && chmod +x /root/scripts/rps_boot.sh
bash -c 'echo 00000f > /sys/class/net/eth0/queues/rx-0/rps_cpus'
bash -c 'echo 0000f0 > /sys/class/net/eth0/queues/rx-1/rps_cpus'
bash -c 'echo 8192 > /sys/class/net/eth0/queues/rx-0/rps_flow_cnt'
#ethtool -G eth0 rx 2047

Network on Windows
• Remove unused network protocols
• Set the correct order of NICs
• Results: well, no big difference

Results from network tuning on Linux
• Much better throughput (users do not claim :)
• Significant decrease of Load Average
• Better distribution of load between CPUs

Conclusion for Linux configuration
• Use server distribution
• Use fresh version (CentOS 7+, Ubuntu Srv 16+)
• xinetd configuration is critical (due to Classic)
• Tune limits for process files, memory, file cache, and network

Conclusion for Windows Tuning
1. Main focus is on RAM tuning
2. CPU tuning is through CPU Affinity restrictions
3. Don’t forget to disable useless services/applications
4. In general Windows has far less parameters to tune, and
they are not clear

Misc Windows Tuning tips
• Enable High Performance Power Plan
• Enable background processes priority
• Disable useless services
• Prefetch/Fetch On/Off – no differences
• Desktop Heap for Classic for non Local System account

Firebird at BudZdorov
• Firebird Classic 2.5
• Why not SuperClassic?
• It is slow for more than 800 connections
• No plans to fix it, since Firebird 3 SuperServer must be used

firebird.conf
• [root@mskv-cbd-new ~]# cat /opt/firebird/firebird.conf
DefaultDbCachePages = 1024
TempCacheLimit = 67108864
TempDirectories = /dev/shm;/3par-vv1/fb_tmp;/tmp
LockHashSlots = 49009
LockMemSize = 82048576
TcpRemoteBufferSize = 1448

TempCacheLimit tips
• Default firebird.conf
• TempBlockSize = 1048576
• May increase to 2 or 3mln bytes, but not to 16mb
• TempCacheLimit = 67108864
• SuperServer and SuperClassic. Classic = 8mb.
• TempDirectories = c:temp;d:temp…
• Increase TempCacheLimit for SuperServer and SuperClassic!

Maintenance and backups
• Automatic sweep is disabled
• All connections are disconnected at 0-00
• Manual sweep is at 00-05
• Verified backup (gbak) – every day at 1am
• Replication works as a standby

Summary for 2.5
• 1500 connections and 453Gb is a acceptable load for the
Firebird 2.5
• Firebird and Linux should be tuned
• Maintenance is the key: sweep, restart of connections,
backups
• Replication is mandatory for protection, since
backup/restore takes 18 hours

Firebird at Customer#2
Firebird 3.0.2
DefaultDbCachePages = 2M
FileCacheSystemThreshold=50M
TempCacheLimit = 9G
LockHashSlots = 21001
LockMemSize = 82048576

Summary
• Firebird 3.0.2 get the biggest benefit from huge number of
page buffer (properly configured)
• Good design (short write) transactions eliminate need for
everyday restarts

Useful links
• Collection of optimized Firebird configuration files
https://ib-aid.com/en/optimized-firebird-configuration/
• Firebird Hardware Guide
https://ib-aid.com/en/articles/firebird-hardware-guide/
• 45 Ways To Speed Up Firebird
• https://ib-aid.com/en/articles/45-ways-to-speed-up-firebird-
database/

Thank you!
• Questions?
• www.ib-aid.com
• ak@ib-aid.com

Tuning Linux Windows and Firebird for Heavy Workload

Recommended

More Related Content

What's hot (14)

Similar to Tuning Linux Windows and Firebird for Heavy Workload (20)

More from Marius Adrian Popa (11)

Recently uploaded (20)

Tuning Linux Windows and Firebird for Heavy Workload