The MMP runs as a single multithreaded process and is CPU and network bound. It uses disk resources only for logging. The MMP scales most efficiently on two processor machines, scales less than linearly from two to four processors and scales poorly beyond four processors. Two processor, rack mounted machines are good candidates for MMPs.
In deployments where you choose to put other component software on the same machine as the MMP (Calendar Server front end, Communications Express web container, LDAP proxy, and so on), look at deploying a larger, four processor SPARC machine. Such a configuration reduces the total number of machines that need to be managed, patched, monitored, and so forth.
MMP sizing is affected by connection rates and transaction rates. POP sizing is fairly straight forward, as POP connections are rarely idle. POP connections connect, do some work, and disconnect. IMAP sizing is more complex, as you need to understand the login rate, the concurrency rate, and the way in which the connections are busy. The MMP is also somewhat affected by connection latency and bandwidth. Thus, in a dial up environment, the MMP will handle a smaller number of concurrent users than in a broadband environment, as the MMP acts as a buffer for data coming from the Message Store to the client.
Never deploy the MMP under HA control. An individual MMP has no static data. In a highly available environment, add one or more additional MMP machines so that if one or more are down there is still sufficient capacity for the peak load. If you are using Sun Fire BladeTM Server hardware, take into account the possibility that an entire Blade rack unit can go down and plan for the appropriate redundancy.
You can put the MMP and Webmail Server on the same set of servers. The advantage to doing so is if a small number of either MMPs or Webmail Servers is required, the amount of extra hardware for redundancy is minimized. The only possible downside to co-locating the MMP and Webmail Server on the same set of servers is that a denial of service attack on one protocol can impact the others.