SpellChecker Load Testing

Revision 07/17/02

Background

When a user accesses the SpellChecker feature on a web site, several requests are issued to the server. These requests are to both the engine and to the web server.

Requests to the engine do the spell-checking work - checking words, getting suggestions for the incorrect words, etc.  These requests begin by invoking the CGI script residing on the server.  This script, in turn, issues requests to the application server that contains the system logic and works with dictionaries.  Each request utilizes its own copy of CGI script, but all these scripts work with only one instance of the application server.  Therefore, such requests impose the load on both web server and application server.

Requests to the web server retrieve static elements, such as images, scripts, style sheets, etc.  Each time the user presses an action button in the SpellChecker dialog, several static items are loaded from the server.

Since the length of text and number of incorrect words vary significantly among users, performance cannot be measured in "sessions", or series of actions from the moment the user started to check text to the moment (s)he submitted the corrected text.  Instead, "request" is the best unit of measurement.  "Request" is defined as the process that occurs when the user presses an action button in SpellChecker dialog window (Change, Ignore). In our testing process, we issued one engine request (specifically, "get all suggestions for the word", as this is most time-consuming action) and four static requests (two .JS files, one .CSS, and one image).

Server Information

The system tested was installed on a server with following parameters:
- Dual AMD 1.53 GHz
- 512M RAM
- Windows 2000 Server
- Apache Web Server

Methodology

To emulate simultaneous user access, a multi-thread stress tool was created.  20 to 30 threads were started on each workstation participating in load testing, with each thread issuing requests to the server. The words to check were drawn randomly from a list of 4,000 correct and 4,000 incorrect words.  These words were the most popular correct and incorrect words, based on analysis of the web system where SpellChecker is used.  The requests contained about 4% incorrect words (this percentage is usual for message boards and online systems).  The number of words was enough to over-run any caches present in the system.

Testing was performed from five workstations, with network connectivity rates ranging from 512K DSL to 100Mbit LAN.

Each stress tool logged its requests.  After testing, the logs from different workstations were combined and the number of requests in each minute was calculated. To achieve this, system clocks were synchronized on all participating workstations.

Results

Testing showed that the SpellChecker server, with the hardware and software configuration described above, was able to process 3500 - 4000 requests per minute.  While the number varied throughout testing, most points were in this range.  System processes on server and workstations and variations in Internet connectivity may explain this variation.

3500 - 4000 requests per minute are equal to:
210 - 240 K requests per hour
5 - 5.5 M requests per day

Testing demonstrated that the only resource critical to overall performance is processor speed.  Memory usage remained stable during the testing and was equal to 80M-90M total (for SpellChecker software, web and ftp servers, operating system etc.).

Processor load on the server oscillated between 20% and 90%, with 45% being the average.

Current configuration of SpellChecker.net

Our system configuration is dual AMD 1.53 GHz, 512M RAM, HDD 40 Gb.