Well, the 'what to buy' discussion fizzled out, and in the meantime Tim Starling did some tests which suggest that the squids are bandwidth-limited, so buying more squids no longer seems like the slam dunk that it did the other day.

I'm leaving for Switzerland in less than 24 hours, but we can continue a longer term discussion of what needs to be bought next, and I can just as easily make an order from there as here.

One concept that came up in IRC that sounds good to me -- the older apache machines accept cheaper non-Ecc non-Registered RAM. It would be relatively inexpensive (say $130/GB) to fill those up with 4Gb of RAM each (which will really only give 3.5GB each usable due to motherboard limitations, so maybe I will just fill them with 3.5GB each to start with).

This RAM could be used for memcached.

We have no really firm data, but Tim believes (reasonably, but he cautions that it is uncertain) that the parser cache hit rate could go from 50% to maybe 85% with enough RAM. A back-of-the-envelope calculation suggests that as much as 50GB could be put to use.

One naturally supposes that the best bang for the buck is at the low end, i.e. taking memcached from 6GB (current, if I'm not mistaken) to say 12GB or 18GB or whatever, would likely result in a significant increase in that hit rate.

My understanding is that increasing the parser cache hit rate is a great way to leverage our apache infrastructure. Spending $1300 for 10 Gb of RAM is likely to do a great deal more good than spending $1050 for another apache.

Comments?

--Jimbo