I tested it with OpenTripplanner (kinda opensource Google transit) on JVM with ARM 8 jre. Times for 380 requests are in seconds:

  ARMv7 Processor rev 2 (v7l) (4) 1.0.0-SNAPSHOT
  min time: 4
  max time: 1655
  mean time: 276.8
  median time: 148.5

Digital ocean 1G memory:

  Intel(R) Xeon(R) CPU E5-2630L v2 @ 2.40GHz (1) 1.0.0- SNAPSHOT
  min time: 1
  max time: 261
  mean time: 51.0447368421
  median time: 30.0

Digital ocean 2G memory:

  Intel(R) Xeon(R) CPU E5-2630L v2 @ 2.40GHz (2) 1.0.0-SNAPSHOT
  min time: 1
  max time: 407
  mean time: 64.6447368421
  median time: 34.0

Local PC:

  Intel(R) Core(TM) i5-3570K CPU @ 3.40GHz (4) 1.0.0-SNAPSHOT
  min time: 0.0001
  max time: 114
  mean time: 21.9026331579
  median time: 12.0

Free redshift cloud (512MB memory):

  Intel(R) Xeon(R) CPU E5-2650 0 @ 2.00GHz (2) 1.0.0-SNAPSHOT
  min time: 5
  max time: 7273
  mean time: 417.507894737
  median time: 103.5

This times doesn't count roundtrip time this is purely JVM time of graph search requests. All requests are serial. All tests except redhift are run with 1GB of JVM memory. Each test was run twice before to warm JVM this is mostly seen in ARM. DO speeds varies a lot. Even bigger VPS doesn't improve speed. ARM speeds doesn't vary a lot but sometimes during testing I didn't get an answer from a server, but this could also be a problem with OTP.

I was impressed with ARM speed (it is much better then free openshift trial maybe because it only has 512MB) but now after free trial I'll probably move to DO the cheapest package, because for 11€ I can get better performance on DO even though it is varied. During normal use I can see the difference in speed between ARM and openshift in requests but not so much between ARM and DO.

Probably ARM would work better when multiple request would come (It has 4 cores) and DO has 1/2.

I don't know if there are any commercial special purpouse ARM JVMs, but HotSpot/OpenJDK is pretty sad performance-wise on ARM.

Oracle provide Java 8 for ARM: http://www.oracle.com/technetwork/java/javase/downloads/jdk8... The performance on my little Raspberry Pi went up dramatically compared to OpenJDK when they released it (over a year ago).

There's a marked absence of benchmark data on the web about this vs same on x86. Does Oracle's license forbid publication of such?

Yeah - was once instructed to use a full container (Glassfish) on ARM. A bootup took 5 minutes, requests took ten seconds. It was depressing.

It would be pretty cheap to hire one of these boxes and try it out.