There is no flame - just a couple of points and a request for data.
If you want to improve something, you should provide solutions,
not critics
Certainly - early next year you will see them. Here are some current performance stats with some new technology we're working on.
Configuration Tool Elapsed Time
(sec’s) Data Transfer Rate
(KB/sec) Requests per Second Requests per
Minute Performance Gain
Factor
Apache Apache Bench 38.735 882.92 2581.64 154,898 1.0
Cyclone Proxy Cache
+
Apache Apache Bench 15.663 2387.79 6384.47 383,068 2.47
Apache Zeus Bench 39.961 855.83 2502.44 150,146.4 1.0
Squid
+
Apache Zeus Bench 28.910 1314.42 3459.01 207,540.6 1.38
Cyclone Proxy Cache
+
Apache Zeus Bench 15.176 2464.42 6589.35 395,361 2.63
Cyclone Proxy Cache
(Tuned Parser)
+
Apache Zeus Bench 13.505 2769.34 7404.67 444,280.2 2.95
Cyclone Proxy Cache (4 Tuned Functions)
+
Apache Zeus Bench 13.006 2875.6 7688.76 461,325.6 3.07
These numbers were obtained using a single processor Itanium® 1.0Ghz (Madison) chip. By tuning certain HTTP string handling functions we have seen up to a factor 11 performance improvement.
Our next benchmark is due by year end. Essentially we will be adding one more line for the stats above. The goal is very simple - transmit greater than 1 million requests in a single minute on a single processor Itanium 1Ghz machine. A factor 10 performance improvement. A single processor Deerfield Itanium® chip costs $744 - our solution doesn't require a current OS, nor hard drive to operate - it scales to multiple chips and can support a cache of up to 1 terabyte of RAM
Revolution is for new players, carefully crafted evolutions are for the
Mass
Yep… Support for a 1TB cache, no hard drive, no current OS required, and the ability to pump data faster than any other platform on the planet should do the trick. Only thing left is to get the Itanium® platform into a single 1RU box at sub $5,000. I doubt we will have to wait long for that.
Long live the revolution
Regards,
Well the http tuning of string handling is a known factor of optimization, just study tomcat 3.2, 3.3 and Coyote 1.1 and you'll that it could still be optimized.
BTW, if you post these benchmarks on the httpd-dev list, should I assume you'll give ASF your optimized&tuned algorythms ?
Do you known that IBM does some nice optimization using FRCA on its Apache 2.0 implementation on iSeries ?
A proof that Apache 2.0 is a great platform for such games ;)
