I was chatting with Richard about server capacity tonight (yeah, we do know how to live it up) and these thoughts came to mind as a result...
It’s been a long time since I had to put real thought into how many traditional Lotus Notes mail client users I could stuff onto a Lotus Domino server. I remember back in the day (Notes 2.1) when the answer was 20 – on an OS/2 1.3 (corrected, it was OS/2 1.3, not 2.1 which came much later and handled a few hundred users) machine. I remember it hitting above 80 in the version 3.0 days and moving into the hundreds with 486 machines running Notes 4.0 servers. Sometime after it moved into the thousands – and now I’m seeing reports of tens of thousands – it stopped being a limitation of the hardware and software.
There is a great deal more to capacity planning than how many users it can handle before the lights go out. Most modern Domino server rollouts are based now on the user population centers, the network layout, and intelligent failover design.
Andrew’s First Rule of Performance is “Availability Comes First”. Unless you like to live dangerously, you have to assume at any given time that one of your machines could go up in smoke—literally. It does happen. For that reason, you need a cluster. Two or more Domino servers running in a cluster makes for a very stable user experience. Most of the time, let the servers do their load balancing, let the users enjoy the performance, and let the hardware work. If one box goes down, users see (at most) a single error screen and are simply using the other box – often without even knowing it.
Performance is also going to be dependant on your network. It’s about both bandwidth – how much stuff can you move between the workstations and servers, and also about latency – how long does it take to move even a small amount. Take a digital satellite link for example. You can get these for not much more than a cable modem link now for your home in the country. They generally offer lots of bandwidth, but very high latency. It just takes a lot of time for the signal to traverse to space and back. The link can feel slow if you’re on a “chatty” network application like Email. For that reason, you need to put your server clusters close to your user populations.
When you consider breaking your users down to a pair of clustered boxes at each of four major networking areas in your company, suddenly, you’re not worried about a Notesbench score of 24,000 users on a $75,000 dollar machine. No, because probably those users are now about 6000 users in each of four sites. Of course, you’d never put 6000 users on one box without failover, so we’re really talking about 3000 users per server – just based on building a smart topology. Now, talking about a cluster to handle 6000 users is a whole lot easier and less expensive than one to handle 24,000 isn’t it.
