Client VM on amd64

[Igor Veresov]

On 2/17/11 1:30 AM, Ismael Juma wrote:
> Paul Hohensee<paul.hohensee at ...>  writes:
>> The reason the commits mention amd64 and compressed oops is because the
>> client compiler is part of the tiered compilation system (c1 + c2)
>> that's built into the server jvm and available via -XX:+TieredCompilation.
>> The 64-bit client compiler must work in order for tiered compilation to work.
>
> What is the status of TieredCompilation in terms of performance and stability?
>
> I've been using it on my desktop (HS20) with IntelliJ IDEA and SBT (build
> system for Scala) for a few weeks now and I've had no crashes as a result
> of it. That is a good sign (in terms of stability, I haven't done any
> performance testing yet), but it would also be good to know if there are
> known issues with it and if there is a plan to solve them.

It should work fine in all configurations (32/64 bit, compressed oops) 
and the performance is good too. Try the latest Oracle's bits for jdk7 
or the upcoming jdk6u25.

This is how it roughly compares to pure C2 with interpreter in specjvm98 
(x86, 32bit):
benchmark: _200_check
startup: 0.052 0.053 diff: -1.92%
settled: 0.0326 0.0336 diff: -3.06%
benchmark: _227_mtrt
startup: 1.822 1.518 diff: 16.68%
settled: 0.4027 0.4023 diff: 0.09%
benchmark: _202_jess
startup: 1.473 1.217 diff: 17.37%
settled: 1.1567 1.0199 diff: 11.82%
benchmark: _201_compress
startup: 4.426 3.905 diff: 11.77%
settled: 4.274 3.8 diff: 11.09%
benchmark: _209_db
startup: 6.562 6.566 diff: -0.06%
settled: 6.2705 6.271 diff: -0.00%
benchmark: _222_mpegaudio
startup: 2.863 2.599 diff: 9.22%
settled: 2.3185 2.2925 diff: 1.12%
benchmark: _228_jack
startup: 2.738 1.581 diff: 42.25%
settled: 1.2824 1.0612 diff: 17.24%
benchmark: _213_javac
startup: 5.602 2.883 diff: 48.53%
settled: 1.6939 1.745 diff: -3.01%

Each subbenchmark is run 10 times. "startup" is the first iteration, 
"settled" is the average of last 5. Less is better (seconds).

igor

>
> Best,
> Ismael
>