NetKernel News Volume 1 Issue 18

March 5th 2010

What's new this week?

• Kernel update
• Some minor fixes and ehancements to core modules.
• First look at NKPerf10 benchmark.
• Portal goes live
• New NK content.
• Imminent blog.

NKSE Repository

kernel: Removed some debug code that got shipped. Fixed a very very very rare race condition potential with async requests. Async requests now get cleaned up if the kernel threads are changed on the fly (this is a bizarre scenario that would never happen in production - but see benchmarking below for how it was discovered).

layer0: Fix to declarative request to handle double square bracket notation properly. If they were malformed you could get NPE.

layer1: Better error message in toDeterminateString transreptor.

module-standard: Enhanced to use initialization hysteresis parameter for commission-cycle resource cache validity. (Sounds obscure, one day remind us to explain how the standard module uses the NK ROC domain to cache transient boot time resource state and thereby linearize the polynomial-time (or worse) cost of determining spacial address space structure).

[All changes (and new packages) are published in the NKSE and NKEE repositories.]

NKPerf10

There's a new toy to try out. The "nkperf" package provides a new control panel tool to measure your platform. It offers a normalized figure of merit for your system. (To get it, synchronize apposite and look for "nkperf" in package list).

In effect it measures the combined quality of your hardware, OS, Java and NK stack. Some relative numbers for you (NKPerf10 mark is the last number)...

AMD Turion™ 64 X2 Mobile Technology TL-60 2GHz: 4.4 Intel® Core™2 Duo CPU T8300 @ 2.40GHz: 7.4 Dual Quad-Core AMD Opteron™ Processor 2352 2GHz (8 cores): 14.9

The details for the 8-core test are:

NKMark10 (alpha): 14.9 Synchronous: 2181 Asynchronous: 2129 Garbage Collection: 884 IO Read: 305 IO Write: 151

For now this is an alpha mark - there are several variables in play and we want to make sure the mark is reflecting real system quality.

For example, all these test systems were 64-bit linux kernel 2.6.31-19 on Java 1.6.18. The 8-core AMD has top sync/async performance but its disk IO read/write is pretty bad (probably RAID related). Whereas the Intel core 2 and the AMD Turion are both laptops with similar IO numbers but the more modern Intel kicks the *rse of the older and slower clocked AMD laptop.

What would be really helpful is if you can give it a try on your systems and send us some data: Hardware, OS, JVM and NKPerf details. Prize for the first one to beat 14.9! (Tip, it may be a good idea to run it back-to-back a few times to give your JVM chance to JIT optimize).

With some more data we can refine the algorithms and put out a definitive standard NKPerf10 tool. Once we define it as standard we'll also set up a service to post anonymized results to and provide a comparison tool in the portal for you to compare your set ups with others.

The NKPerf10 figure is kind of handy (think of it as a taster) but we have some really useful system tuning/analysis tools which we plan to get out next week. These provide answers to questions like "What's my optimal number of kernel threads on this N-core CPU?", "Would it be a good idea to deploy a throttle in my architecture?" etc, etc.

[Incidentally the emergence of these tools explains why you've been getting kernel updates recently - they create very extreme kernel loads which has allowed us to find and smooth out some really obscure corner cases. For example, in some tools we're now dynamically changing the number of kernel threads - hence the earlier kernel update which fixes the potential of hanging asnyc requests when the number of threads are altered mid-request]

Other News

The NetKernel services portal is now live and integrated with our main website (it can be reached as first item in the top right menu list on any page). Thanks again to the pre-release testers.

We also updated the main Netkernel page (this is where most people arrive when they're looking for NK). We would appreciate your feedback on the content...

http://www.1060research.com/netkernel/

The online version of the reference documentation is also now linked in off the top menu.

Blog

There is definitive action on the blog front...

http://durablescope.blogspot.com/

(If its not published when you get this, try later... Mr B said something about having to go and watch a football match first!).

Have a great weekend.