Hi, I have had a few questions about what we have done to address the server problem that we discovered during our first live test of the new Web site. Here an analysis, written by Jon Bloom, of what actions he has taken. If you want further information, Jon has written a more detailed, very technical document on this topic that I can send you. I am also distributing this to the Board, Fathom and additional in-house staff for their information. 73, Harold Harold Kramer, WJ1B Chief Operating Officer ARRL - The national association for Amateur Radio(tm) 860 594 0220 hkramer@arrl.org ------------------------------------------------------------------------ ------------------------- From: Jon: Our initial attempts and additional testing revealed that the server we were using was inadequate to the task of serving up the pages of the new Web site. A simplified explanation is that each page viewed on the new site needs about 10 times as much memory and 10 times as much processor power to generate compared to a page of the old site. This drastically limited the number of pages that could be served simultaneously, even though the new site was running on a server about twice as capable as the one serving the old site. In addition, using a single server queue for all requests meant that when page generation slowed down, requests for image files (which take no processor power to speak of) often waited in the same queue. The result was that users experienced dreadfully slow page loads. To address this problem, I spent much of last week devising a new server architecture. This architecture is based on a cluster of "cloud" servers. It splits the generation of pages across a number of servers, allowing the generation of far more pages simultaneously. In addition, images and other static files are served from a queue capable of handling many more simultaneous requests. The new architecture provides more scalability, too. If page generation becomes a limiting factor, it is a simple matter to add additional servers to generate more pages in parallel. Looking at two days of traffic on the old site, we see a busiest hour of 5,141 page views and 74,015 total hits. The busiest minute was 187 page views and 1736 total hits. The new cluster was load-tested and measured to serve approximately 1,000 page views and 14,000 total hits per minute, about five times the load of the busiest minute seen during those two days. (Since that load test was performed, additional page caching has been added to the page-generation code that should further improve performance.) Jon Bloom, KE3Z IT Manager ARRL - The national organization for Amateur Radio 225 Main St Newington, CT 06109 860-594-0276 jbloom@arrl.org <mailto:jbloom@arrl.org>