PerformanceNYC
as of mid-November 2012 (commit f5685d20d4124310b4cb3fb7f1934a92e7606bb7)
These measurements were made on a New York City graph including 11 GTFS feeds and OSM data for the entire transit-served area. Using our profiling tools, we queried an OTP server with 10500 requests generated by combining a range of routing parameters (time of day, maximum walk distance, transporation modes) with endpoints chosen randomly but located within 2km of a transit stop.
[
]
This server was configured to use the retrying path service backed by the A* algorithm. Server heap space was set to 6GiB. An identical run was performed on two different EC2 instance types for comparison.
7.5 GiB memory and 2 cores (in this case, 2 non-hyperthreaded cores of a shared 4-core E5507)
Amazon rates this instace at 2.5 EC2 units per core.
17GiB RAM and 2 cores (in this case, 2 hyperthreaded cores on a shared 4-core X5550)
Java heap space was intentionally limited to 6GiB, so the main difference here is that the processor is rated at 3.25EC2 units per core.
In a public-facing server or behind a load balancer, you may want to use EC2 instances with similar computing capacity per core to the above test cases, but with more cores to handle more simultaneous requests. In the 2.5 EC2 unit / core category you might choose a c1.xlarge instance with 8 cores. In the 3.5 EC2 unit / core category you might choose an m3.xlarge instance with 4 cores.
[
]