On Tuesday, January 29, 2013 7:49:18 PM UTC-8, Bik Dhaliwal wrote:
>
>
> Out group at my company is considering using the Zing JVM from Azul,
> it promises zero GC pauses and in certain cases can vastly increase
> performance and lower latency.
>
> I was wondering has had any experience dealing with this product and the
> company.
>

I'm obviously biased (I'm Azul's CTO), but I just wanted to correct the 
notion that we promise "zero GC pauses" in the above. I don't like 
promising "Zero GC pause" notion.

What we do is bring pause times down to well below OS scheduling noise. 
While we completely decimate GC compared to there JVMs (by 100x-1000x 
factors at times), technically there are still 4 pauses per GC cycle. Those 
pauses are VERY short (as in milliseconds or less), and mostly deal with 
phase-shifting the collectors. Specifically, unlike other collectors and 
JVMs, Zing's pauses do not do any of the actual bulk GC work items like 
marking, copying or compacting, weak/soft/phantom/final ref processing, 
etc. All that work happens concurrently and outside of pauses and 
safepoints, making Zing's pause behavior not just super-short, but also 
independent of things like heap size, shape, and allocation rate. 

To use a picture instead of words, below is a sample graph showing all 
pauses (both GC and other safepoints like deoptimazation) from a busy 
multi-hour run of a heavy e-commerce workload at an actual (anonymized) 
customer [the longest pause seen in this case is around 5msec]. This same 
workload shows regular multi-second oldgen pauses and tens to hundreds of 
msec of newgen pauses on "well tuned" Oracle and OpenJDK JVMs.  For Zing, 
most of the "longer" (above 1-2 msec) stuff has to do with scheduling noise 
and cpu contention. In finely tuned low latency environment we see stuff 
that goes well below that.

<https://lh5.googleusercontent.com/-GIGWJ1iUBI4/UQ3W8tvN_vI/AAAAAAAAADY/KgxbevgNDak/s1600/PastedGraphic-3.a.png>
  

-- 
You received this message because you are subscribed to the Google Groups "Java 
Posse" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To post to this group, send email to [email protected].
Visit this group at http://groups.google.com/group/javaposse?hl=en.
For more options, visit https://groups.google.com/groups/opt_out.


Reply via email to