Coyote Web Server

Coyote Web Server Benchmarks

July 15, 1998
Updated September 5, 1999

A series of benchmarks has just been completed with the PicLan-IP Coyote Web Server. This paper discusses the results of these benchmarks.

What this Benchmark Tries to Test

This benchmark is of a single, very simple dynamic page. The page has a single mv/BASIC insertion point and the absolute minimum of user mv/BASIC code (a single TIMEDATE() function) strictly to demonstate and test the overhead of getting to an active content page.

This benchmark is meant to show the performance of PicLan-IP and Coyote and not to test your application code's impact on performance. When looking at these numbers, remember that this is the baseline for the web server's performance and your code will subtract from these numbers.

Benchmark HTML Content

The tests were run against a single HTML page that contained a single, and very simple, dynamic HTML page. A total of nine runs were made with this page ranging from an extremely small 178 bytes up through a size of 204178 bytes. All of the page sizes had a single TIMEDATE() inserted at the top of the page. The rest of the page's contents were static.

The RAW Benchmark Numbers

Page size including HTTP headers	Hits/sec	Bytes/sec
317	84.18	26,684
1,337	81.04	108,347
2,357	78.62	185,299
5,417	70.13	379,874
10,517	59.24	623,045
20,717	43.18	894,516
51,317	22.52	1,155,788
102,317	13.09	1,339,228
204,317	6.49	1,326,734

Conclusion

The benchmark behaved about the way that it was supposed to.

As the page size grows, the hits/second goes down and the bytes/second goes up.
T-1 line speeds were reached with page sizes of under 2500 bytes

Most non graphics intensive web sites deal with average page sizes of 4K-5K
Large HTML forms based sites usually have 5K-15K average page sizes
Graphic intensive web sites often have much larger page sizes

Full 10 Megabite Ethernet was saturated at about 25K page sizes
Above 100K page size, the string manipulation in mv*Base started to degrade a little

This is apparent in the bytes/second for 200K pages being slightly lower than or 100K pages.

Additional tests with static-content pages showed:

Static pages are roughly 10% to 15% faster than same-size dynamic pages

This test quite conclusively shows that PicLan-IP and Coyote can be used for high-volume web sites using full T-1 lines or greater. It should be noted that the test system was far from exotic in terms of hardware configuration and CPU performance (the system had a hardware cost to build of under $1350).

Using the 2.3K page size as an average (which is lower than most sites), Coyote can produce:

75+ hits per second more than saturing a full T-1 internet connection

This translates to:

4500+ hits/minute
250,000+ hits/hour
2,160,000+ hits/8-hour day
5.3 Gigabytes/8-hour day
16.0 Gigabytes/24-hour day
160 Gigabytes/month (assuming usage averaging 8 hours)
480 Gigabytes/month (assuming 24-hour a day saturation)

We believe that this allows Coyote to drive most large MultiValue internet sites without requiring additional web server acceleration techniques and still leaving plenty of processing power for non-Web MultiValue applications.

The Benchmark Configuration Details

Server System	x86 based PC with/ Dual Pentium-200 MMX processors TYAN Tomcat-4D Motherboard 512K Pipelined level-two cache Up to 512Meg of main memory cachable 128Meg of EDO DRAM 6.4Gig IDE disk drive 5400 RPM IBM drive TYAN DMA/33 IDE driver Windows NT 4.0 Server Service Pack 3 installed Base services only running Fast Ethernet Network Interface Bay Network Netgear 10/100 PCI Ethernet Adapter uses Digital 21140 chipset
MultiValue Host	mv*Base build 1.2.01
PicLan-IP Release	2.0.0.116
PicLan-IP Configuration	One IP address SMPT/POP3 Mail services were running
Coyote Web Server Configuration	Standard Configuration Logging was enabled The log file was cleared each test run The log file was sized to prevent it from going into overflow Content cache checking was enabled and set at 30 seconds Four phantom processes were used to process web requests This is the recommended number for a dual-proc system
Client Test System	x86 based PC with/ Single Cyrix PR200+ CPU (150MHz core) 512 K Pipelined level-two cache Motherboard only caches first 64Meg of system memory 128Meg of Fast Page Mode DRAM Windows NT Server 4.0 Service Pack 3 installed Base services plus DNS and IIS servers running Fast Ethernet Network Interface Compex 10/100 PCI Ethernet Adapter uses Digital 21143 chipset
Client Test Software	WebStone for Windows NT Client helper only. No supervisor used Cnfigured for 5 threads Non-formal tests were run with 5 to 50 threads No signifigant performance difference was noted No connection errors were encountered on any tests at any thread count

Mini Benchmarks with D3/Linux new

Another set of benchmarks was conducted with D3/Linux running on dual-processor 400MHz Linux systems over fast ethernet. With small active-content pages, the system satured at about 212 page hits/second or over 750,000 hits/hour. Even assuming that your application degrades page performance down to 100 hits/second, at 10K pages this is a sustained rate of 1000K bytes/second which is full ethernet speeds or between 6 and 7 dedicated T-1 lines. All with pure dynamic content.

When tested with large static pages, testing actually saturated fast-ethernet at about 85 Megabits/second so no "real" numbers are being quoted. Regardless, most users will simply not experience server performance problems.