30GB innodb_buffer_pool_size?

[James Day]

For Wikipedia's English database servers, which make up about half of all database load, I'm considering recommending the purchase of quad Opteron servers with 16GB or 32GB of RAM (and 12-16 or more 15K SCSI drives).

What can I/we expect to see in the way of issues with buffer pools in this size range? How has testing and production use gone? Are there a few customers using equipment of the 16-32GB capacity in production, ideally in a Linux/Opteron setting? What's the theoretical maximum capacity and how high have you gone in practical testing?

As well as read caching, one of my objectives at this size is to hold a large portion of the write working set in dirty pages in the buffer pool, to reduce disk write loads. Here, the 4GB maximum combined log file size is an issue. How many GB of dirty pages can I expect to hold? 2GB? 4GB?

In theory I'm not interested in such large servers. I'd much prefer smaller. But I have to anticipate rapid doubling rates and delivery of software changes allowing load spreading not happening fast enough. Cluster and maybe the federated engine are definitely of interest... but may not be suitably stable and able to handle our record lengths and capacity needs before we need the capacity, so we need to be planning and buying for a few doublings today without assuming their availability. Doesn't stop me from wanting disk-capable Cluster with a year of integration work to be available yesterday though...:)