Last year at HP we undertook a programme of performance measurements on Swift running on a variety of hardware. We measured the CPU burden and I/O rates of individual Swift processes under a variety of conditions, ranging from idle (i.e. no external PUT/GET activity) to very high transaction loads. We used the getput tool to impose loads on our test hardware, and we concentrated our measurements on PUTs of small objects of 1KB and 4KB. We used collectl to gather performance data. Some of the hardware that we tested is currently in use in our production public Swift, and some of it is hardware that we plan to put into production. We were particularly interested in optimising the balance between CPU, RAM and I/O for component Swift services (proxy, object and container/account). In this presentation we will summarise the performance measurements themselves, outline our conclusions and show that with the appropriate hardware you can realise significant improvements in Swift transaction rates.
