Authors Jim Jagielski
License CC-BY-3.0
Apache httpd v2.4:
Watch Out Cloud
-or-
Hello Cloud: Buy you a drink?
Jim Jagielski
About me
§ Jim Jagielski
- Hacker and developer
- Co-founder of the ASF
- Member, Director and President
- Director: Outercurve and OSI
- Council member: MARSEC-XL
- Consulting Engineer with Red Hat
- @jimjag
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
What we will cover
§ Overview of Apache httpd 2.4
- General improvements
- Reverse proxy improvements
§ How the Cloud is a game-changer for web
§ Performance Related Enhancements
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Apache httpd 2.4
§ Currently at version 2.4.2 (2.4.1 went GA Feb 21, 2012)
§ 2.4.3 RSN
§ Significant Improvements
- high-performance
- cloud suitability
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Apache httpd 2.4 - design drivers
§ Support for async I/O w/o dropping support for older systems
§ Larger selection of usable MPMs: added Event, Simple, etc...
§ Leverage higher-performant versions of APR
§ Increase performance
§ Reduce memory utilization
§ The Cloud
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
What’s New: Apache httpd 2.4
§ Bandwidth control now standard
- mod_ratelimit
§ Finer control of timeouts, esp. during requests
- mod_reqtimeout
- KeepAliveTimout down to the millisecond
§ Finer control over logging
- per module/per directory
- new logging levels (TRACE[1-8])
§ <If> supports per-request conditions
§ slot-based shared memory capability
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
What’s New: Apache httpd 2.4
§ Controllable buffering of I/O
- mod_buffer
§ Support for Lua (still experimental as of 2.4.2)
§ Loadable MPMs
§ General purpose Response Body substitution
- mod_sed
§ Auto-convert Response -> RFC 2397 data URL
- mod_data
§ Config file variables
§ Cache improvements
§ Proxy improvements (‘natch)
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Why Proxy Matters
§ Cloud puts big focus on horizontal scaling
§ Apache httpd still the most frequently used front-end
§ Proxy capabilities must be cloud friendly
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Proxy Design Drivers
§ Becoming a robust but generic proxy implementation
§ Support various protocols
- HTTP, HTTPS, CONNECT, FTP
- AJP, FastCGI, SCGI, WSGI (soon)
- Load balancing
§ Clustering, failover
§ Performance
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
What’s New: Apache httpd 2.4 proxy
§ Reverse Proxy Improvements
- Supports FastCGI, SCGI in balancer
- Additional load balancing mechanisms
- Runtime changing of clusters w/o restarts
- Support for dynamic configuration
- mod_proxy_express
- mod_proxy_html
- mod_fcgid
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Load Balancer
§ mod_proxy_balancer.so
§ mod_proxy can do native load balancing
- weight by actual requests
- weight by traffic
- weight by busyness
- lbfactors
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Load Balancer
§ Backend connection pooling
§ Available for named workers:
- eg: ProxyPass /foo http://bar.example.com
§ Reusable connection to origin
- For threaded MPMs, can adjust size of pool (min, max, smax)
- For prefork: singleton
§ Shared data held in shared memory
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Load Balancer
§ Sticky session support
- aka “session affinity”
§ Cookie based
- stickysession=PHPSESSID
- stickysession=JSESSIONID
§ Natively easy with Tomcat
§ May require more setup for “simple” HTTP proxying
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Load Balancer
§ Cluster set with failover
§ Group backend servers as numbered sets
- balancer will try lower-valued sets first
- If no workers are available, will try next set
§ Hot standby
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Putting it all together
<Proxy balancer://foo>
BalancerMember http://php1:8080/ loadfactor=1
BalancerMember http://php2:8080/ loadfactor=4
BalancerMember http://phpbkup:8080/ loadfactor=1 status=+h
BalancerMember http://phpexp:8080/ lbset=1
ProxySet lbmethod=bytraffic
</Proxy>
<Proxy balancer://javaapps>
BalancerMember ajp://tc1:8089/ loadfactor=1
BalancerMember ajp://tc2:8089/ loadfactor=4
ProxySet lbmethod=byrequests
</Proxy>
ProxyPass /apps/ balancer://foo/
ProxyPassReverse /apps/ balancer://foo/
ProxyPass /serv/ balancer://javaapps/
ProxyPass /images/ http://images:8080/
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Embedded Admin
§ Allows for real-time
- Monitoring of stats for each worker
- Adjustment of worker params
- lbset
- load factor
- route
- enabled / disabled
- ...
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Embedded Admin
§ Allows for real-time
- Addition of new workers/nodes
- Change of LB methods
- Can be persistent
- More RESTful
- Can be CLI-driven
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Easy setup
<Location /balancer-manager>
SetHandler balancer-manager
Order Deny,Allow
Deny from all
Allow from 192.168.2.22
</Location>
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
Click here
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
Click here
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
Changing the
LBmethod
Adding new worker
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Admin
Wow!
Wow!
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Mass Reverse Proxy
§ Use the new mod_proxy_express module
- ProxyPass mapping obtained via db file
- Fast and efficient
- Still dynamic, with no config changes required
ProxyExpress map file
##
##express-map.txt:
##
www1.example.com http://192.168.002.2:8080
www2.example.com http://192.168.002.12:8088
www3.example.com http://192.168.002.10
...
www6341.example.com http://192.168.211.26
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
What’s on the horizon?
§ Improving AJP
§ Adding additional protocols
§ More dynamic configuration
- Adding balancers!
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Cloud and Performance
§ The Cloud is a game changer for web servers
- Horizontal scalability is no longer as painful
- Concurrency is no longer the sole consideration
- ... or even the primary one
- What’s important now? Transaction Time!
- Low latency
- Fast req/resp turnover
- Does density still matter? Of course!
- Are there environs where concurrency is the bugaboo? You betcha! (but the cloud makes
these more and more rare)
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Apache httpd vs nginx
§ Why nginx? Everyone asks about it...
§ Benchmark: local and reverse proxy transaction times
- Apache httpd 2.4.1-dev, nginx 1.2.0
- Fedora 16, Dual Xeon 2.28GHz
- 4GB memory
- localhost loopback and external (no firewall)
- Double checked results: OSX, Ubuntu 10.04
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Setup
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Setup
Setup 1: loopback
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Setup
Setup 1: loopback
Setup 2:
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Setup
Setup 1: loopback
Setup 2: Setup 3:
Setup 3:
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Considerations
§ Multiple benchmarking systems:
- flood (50/250/5/2, 50/100/5/2, 50/5/5/2)
- httperf (num-conns=100->10000, numcalls=3)
§ Full URL requests (www.example.com/index.html)
§ Static local requests
§ Static reverse proxy requests
§ All Apache httpd MPMs
§ No significant “tuning” efforts (mostly out of the box configs)
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
nginx vs Event (typical)
nginx Apache - Event MPM
2,000 2000
1,500 1500
1,000 1000
500 500
0 0
Open Write Read Close
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
nginx vs Worker (typical)
nginx Apache - Worker MPM
2,000 2000
1,500 1500
1,000 1000
500 500
0 0
Open Write Read Close
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
nginx vs Prefork (typical)
nginx Apache - Prefork MPM
2,000 2000
1,500 1500
1,000 1000
500 500
0 0
Open Write Read Close
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Focus on open()
Comparison - opens
2000
1500
1000
500
0
Prefork Worker Event nginx
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Focus on write()
Comparison - writes
2000
1500
1000
500
0
Prefork Worker Event nginx
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Focus on read()
Comparison - reads
2000
1500
1000
500
0
Prefork Worker Event nginx
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Total req/resp time
Comparison - total transaction (close)
2000
1500
1000
500
0
Prefork Worker Event nginx
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Resp to Req. Bursts - httpref
100 ---> 10000
7.00
5.25
3.50
1.75
0
min avg max dev min avg max dev min avg max dev min avg max dev min avg max dev min avg max dev
prefork worker event nginx
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Independent benchmark
#!/bin/sh
RESULT='./result.txt'
for port in 80 8080 8888
do
#for count in 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
#for count in 11000 12000 13000 14000 15000 16000 17000 18000 19000 20000
for count in 21000 22000 23000 24000 25000 26000 27000 28000 29000 30000
do
echo -n "$port $count " >> $RESULT
httperf --rate $count --num-conns 25000 --server ipaddr --port $port \
--uri=/test.html | grep "Request rate:" >> $RESULT.$port
sleep 60
done
done
Source: Ryosuke Matsumoto : http://blog.matsumoto-r.jp/?p=1812
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Benchmark Conclusions
§ Events, polling and fork/spawn creates overhead: good for “more bang for buck” system,
bad for performance for that request
§ For concurrency, Event & Worker on par with nginx*
§ For transaction speed, prefork shines
§ Let’s reboot “Simple” mpm (currently being done)
§ *Main Caveats:
- Apache is never resource starved
- If memory is a scarce resource, nginx still better (for now ;) )
- More work can (and should) be done
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
In conclusion...
§ Performance of Apache httpd 2.4 still in the big leagues (and on par with the “big boys” and
the fanboi webserver du jure)
§ For cloud environs, the performance and dynamic control of Apache httpd 2.4 in reverse
proxies is just what the Dr. ordered (and flexibility remains a big strength)
§ Architecture of Apache httpd 2.4 allows a lot of room for growth and additional functionality
(both for the cloud and not)
§ There’s still a category of “edge cases” that require nginx, lighttpd, G-WAN, Apache Traffic
Server, etc... If that’s you, don’t try to use Apache httpd (but if you do, provide patches!)
§ lies, damned lies and benchmarks (sorry, statistics).
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
Thanks!
§ Contact Info:
- Jim Jagielski
- jim@jaguNET.com jimjag@redhat.com
- @jimjag www.jimjag.com
- people.apache.org/~jim/presos/
This work is licensed under a Creative Commons Attribution 3.0 Unported License.