Things to consider when optimizing at the 99th percentile. Variance is key and is one of the biggest reasons EBS blows.

Counterintuitively, he suggests synchronizing variances across a cluster (for example, a Puppet run). Even though for that moment in time all requests are slow, it's better to do that than let every request potentially be slow because it fanned-out to a temporarily slow node.

Interesting call graph visualization.

Compression library designed for the real last mile: CPU cache lines.

SimpleGeo's application-level monitoring and introspection tool.

Artur says this will tell you what's slow.

DTrace for Linux. Ish.

Please to email r@rcrowley.org if you know what the fuck lcnum and bnum actually mean in tuning a Tokyo Cabinet B+Tree database.

"If you can afford the loss of some of the latest committed transactions if a crash occurs, you can set the innodb_flush_log_at_trx_commit parameter to 0. InnoDB tries to flush the log once per second anyway, although the flush is not guaranteed."

Oh lord, help me.

David installed irqbalance on a few of my machines. Seems to promise to fix some Linux bit-rot.

System wide profiling. Now I know exactly how much of my time is spent in red-black trees.