Scaling PHP/MySQL

Hyatt Regency San Francisco Airport Burlingame, CA

San Francisco, CA

October 18-21, 2005

Hardware Layouts for LAMP Installations

John Allspaw, Flickr PlumbrFlickr (Yahoo)

[email protected]

October 18, 2005

Oct. 18, 2005 # 2


Hardware requirements for LAMP installs have to do with:

o A decent amount about the actual hardware (“in-box” stuff)

o A bit more about the hardware architecture

o Which should complement the application architecture

Oct. 18, 2005 # 3


What we’ll talk about here:

o Database (MySQL) layouts and considerations

o Some miscellaneous/esoteric stuff (lessons learned)

o Caching content and considerations

Oct. 18, 2005 # 4


• Growing Up, “One Box” solution

Basic web application (discussion board, etc.) Low traffic Apache/PHP/MySQL on one machine Bottlenecks will start showing up:

• Most likely database before apache/php• Disk I/O (Innodb) or locking wait states (MyISAM)• Context switching between memory work (apache) and

CPU work (MySQL)

Oct. 18, 2005 # 5


ONE BOX

Oct. 18, 2005 # 6


• Growing Up, “Two Box” solution

Higher traffic application (more demand) Apache/PHP on box A, MySQL on box B Same network = bad (*or is it ?), separate network =

good Bottlenecks with start to be:

• Disk I/O on MySQL machine (Innodb)• Locking on MyISAM tables• Network I/O

Oct. 18, 2005 # 7


TWO BOX

Oct. 18, 2005 # 8


• Growing Up, “Many Boxes with Replication” solution

• Yet even higher traffic • Writes are separated from reads (master gets

IN/UP/DEL, slaves get SELECTs)• Diminishes network bottlenecks, disk I/O, and

other “in-box” issues• SELECTs, IN/UP/DEL can be specified within the

application, • OR….• Load-balancing can be used

Oct. 18, 2005 # 9

Hardware Layouts for LAMP InstallationsMANY BOX

Oct. 18, 2005 # 10


Slave Lag

• When slaves can’t keep up with replication• They’re too busy:

• Reading (production traffic)• Writing (replication)

• Manifests as:• Comments/photos/any user-entered data doesn’t

show up on the site right away• So users will repeat the action, thinking that it

didn’t “take” the first time, makes situation worse

Oct. 18, 2005 # 11


Insert funny photo here about slave lag*

*slave lag isn’t funny

Oct. 18, 2005 # 12


Hardware Load Balancing MySQL

Oct. 18, 2005 # 13


How It’s Usually Done

• Standard MySQL master/slave replication

• All writes (inserts/updates/deletes) from application go to Master

• All reads (selects) from application go to a load-balanced VIP (virtual IP) spreading out load across all slaves

Oct. 18, 2005 # 14


Oct. 18, 2005 # 15


What Is Good About Load Balancing

• you can add/remove slaves without affecting application, since queries are atomic (sorta/kinda)

• additional monitoring point and some automatic failure handling

• you can treat all of your slave pool as one resource, and makes capacity planning a lot easier if you know the ceiling of each slave

Oct. 18, 2005 # 16


• How do you know the ceiling (maximum QPS capacity) of each slave ?

• First make a guess based on benchmarking (or look up some bench results from Tom’s Hardware or anandtech.com, etc.

• Then get more machines than that :)

• Scary: in production during a lull in traffic, remove machines from the pool until you detect lag

• The QPS you saw right before slave lag set in:THAT is your ceiling

Oct. 18, 2005 # 17


Oct. 18, 2005 # 18


What Can Be Bad/Tough About Load Balancing:

• not all load-balancers are created equal, not all load-balancing companies expect this product use, so support may still be thin

• not that many people are doing it in high-volume situations yet, so support from community isn’t large either

• Gotchas: • port exhaustion, • health checks, • and balance algorithms

Oct. 18, 2005 # 19


Port ExhaustionPROBLEM:

• LB is basically a traffic cop, nothing more

• Side effect of having a lot of connections: only ~64,511 ports per each IP (VIP) to use

• 64,511 ports/120 sec per port….

• ~535 max concurrent connections per IP*

* Not really, but close to it: tcp_tw_recycle and tcp_tw_reuse

Oct. 18, 2005 # 20


Oct. 18, 2005 # 21


Port Exhaustion (cont’d)SOLUTION:

• Use a pool of IPs on the database slave/farm side (Netscaler calls these “subnet IPs”, Alteon calls them “PiPs”)

• Monitor port/connection usage, know when it’s time to add more

Oct. 18, 2005 # 22


Health checks

• LB won’t know anything about how well each MySQL slave is doing, and will pass traffic as long as port 3306 is answering

• Load balancers don’t talk SQL, only things like plain old TCP, HTTP/S, maybe FTP

Oct. 18, 2005 # 23


Health checks (cont’d)• Two options:

1. Dirty, but workable: Have each server monitor itself, and shut

off/firewall its own port 3306, even if MySQL is still running

Oct. 18, 2005 # 24


Health checks (cont’d)

2. Cleaner, but a bit more work:• Have each server monitor itself, and run a

check via xinetd (for example, a nagios monitor)

So the LB can tickle that port, and expect back an “OK” string. If not, it’ll automatically take that server out of the pool

Good for detecting and counteracting isolated incidents of ‘slave lag’ and automatically handling it

Oct. 18, 2005 # 25


Health Checks

Oct. 18, 2005 # 26


Balancing Algorithms

• Load balancers know HTTP, FTP, basic TCP, but not SQL• Two things to care about:

• Should the server still be in the pool ? (health checks)

• How should load get balanced ? “least connections” or “least bandwidth” or

“least anything” = BAD Because not all SQL queries are created equal Use “round-robin” or “random” What happens if you don’t: Evil Favoritism™

Oct. 18, 2005 # 27


Evil Favoritism

Oct. 18, 2005 # 28


Oct. 18, 2005 # 29


• Meanwhile….for “in-the-box considerations”

Interleaving memory *does* make a difference

Always RAID10 (or RAID0 if you’re crazy*) but NEVER RAID5 (for Innodb, anyway)

RAID10 has much more read capacity, and a write penalty, but not as much as RAID5

Always have battery backup for HW RAID write caching

Or, don’t use write caching at all

Oct. 18, 2005 # 30


• “IN-THE-BOX” considerations (cont’d)

Always have proper monitoring (nagios, etc.) for failed/rebuilding drives

SATA or SCSI ? SCSI ! It’s worth it! 10k or 15k RPM SCSI ? 15k! It’s worth it!(~20% performance increase when you’re disk bound)

For 64bit Linux (AMD64 or EM64T):• Crank up the RAM for Innodb’s buffer pool• Swapping = very very bad either:

• Turn it off (slightly scary)• Leave it on and set /proc/sys/vm/swapiness = 0

Oct. 18, 2005 # 31


• 10k versus 15k drives ?

• Does it really matter that much ?

• Some in-the-wild proof….

Oct. 18, 2005 # 32


10K drives

15K drives

Slave Lag in production

Oct. 18, 2005 # 33


• Using MySQL with a SAN (Storage Area Network)

• Do layout storage same as if they would be local• Do make sure that the HBA (fiber card) driver is well

supported by Linux• Don’t share volumes across databases• Don’t forget to correctly tune Queue Depth Size,

which should be increasing, from server HBA -> switch -> storage

Oct. 18, 2005 # 34


Caching your static content

Oct. 18, 2005 # 35


• Caching Static Content

SQUID = good Relieve your front-end PHP machines from looking up

data that will never (or rarely) change Generate static pages, and cache them in squid,

along with your images

Oct. 18, 2005 # 36


• Caching Static Content (cont’d)

Use SQUID to accelerate plain-old origin webservers, also known as “reverse-proxy” HTTP acceleration

Described here and elsewhere: http://www.squid-cache.org/Doc/FAQ/FAQ-20.html

Oct. 18, 2005 # 37


Basic SQUID layout

• squid accepts requests on 80• passes on cache misses to apache on 81• apache uses as its docroot an NFS mounted dir• should be on local subnet, or dedicated net

Oct. 18, 2005 # 38


• Good HW layout for high-volume SQUIDing

Do use SCSI, and many spindles for disk cache dirs Don’t use RAID Do use network attached storage, or place the origin

servers on separate machines Do use ext3 with noatime for disk cache dirs Do monitor squid stats

Oct. 18, 2005 # 39


Flickr: How We Roll

Oct. 18, 2005 # 40


• Yummy SQUID stats:

• >2800 images/sec, ~75-80% are cache hits

• ~10 million photos cached at any time• 1.5 million cached in memory

Oct. 18, 2005 # 41


The End

Scaling PHP/MySQL

Documents

lamp installationsslave lag

health checks

lamp installations

lamp installationswhat

hardware layouts

server monitor

created equal

load balancing