This page describes some of the measurements we have doneduringdevelopment of MirageOS, and descriptions of some of the problems we encountered and how they were fixed. We use DNS serving performance as the first measure of I/O performance, for several reasons:
UDP is more straightforward to implement than TCP and is stateless. This lets us test the performance of the event loop, Xen rings and the Ethernet/IPv4 stack without getting mixed up in the intricacies of TCP at the same time.
The small size of most DNS packets is a pessimal test of Xen PV IO (grant tables require that a 4K page must be allocated per packet, most of which is unused), and any OCaml-related slowdowns in the data path will become rapidly obvious. The intuition is that if DNS performs well, then "big data" manipulation should be good too.
The size of the record set being looked up lets us smoothly move from being I/O bound (with a small number of records) to CPU-bound (with a large number of records to look up per request). We have spotted performance cliffs here related to garbage collection, and experimented with memoization techniques to smooth things out.
The queryperf test suite provides realistic test data that we can also run against other servers.
BIND and NSD are widely used and critical servers that we can use as a benchmark, and aim to beat! BIND is an easy target, but NSD is supremely well tuned for high performance.
All of the DNS tests are in mirage-perf.git/dns repository, and can be run on a Debian or Ubuntu Xen box. We have used Xen 3.4, 4.0 and 4.1. It should work with xm but we mostly use the new xl command line tool.
The tests currently perform the following tasks:
bin/setup.shuses debootstrap to set up domU images of Linux. (and install NSD/BIND?)
bin/generate.sh runs queryperf to create the test zone files, and compiles the MirageOS UNIX and kernel binaries for different result sizes
bin/run.sh reads some parameters (ranges of record sizes to test, which servers to run), and launches a series of client and server domains that run queryperf and responses respectively. The packets go across a dom0 bridge with all hardware checksumming disabled. Queries per second and response time variances are logged into dns/data/
If you do ever decide to run these tests and try to reproduce our results, please do archive up your dns/data/directory and send them to us too, along with a description of the hardware and Xen/Linux versions you used.
Stay tuned, still working on publishing these...