User: Password:
|
|
Subscribe / Log in / New account

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 14, 2006 13:53 UTC (Thu) by richardm (guest, #7750)
Parent article: KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Bryan Cantrill and his cohorts are certainly to be congratulated in raising awareness for this type of technology. But whether it can be claimed that they have earned a place in the "Kernel Hacker's Bookshelf for describing the first system that lives up to this ideal" is arguable. Preceding dtrace is a lineage of evolving technologies that go back to the dawn of commercial mainframes. Their paper references the immediately preceding work on Dynamic Probes for Linux, which was published 2001 at FREENIX. That work acknowledges its immediate predecessor, DTRACE for OS/2 (ca 1994) - yes that's right the name isn't even original. OS/2's DTRACE provided a comprehensive low-level scripting language that allow system-wide instrumentation to be applied to a running system. And that language permitted data gathering, rudimentary statistics and triggers for other debugging capabilities. DTRACE arose from an earlier implementation in OS/2 Version 1, the user interface to which was provided by a high-level scripting language and a language interpreter called TRCUST. Besides providing dynamic tracing, TRCUST also provided dynamic profiling. All of that has its origins in similar technologies developed in the 1970s. An example being the Dynamic Support System, which was part of IBM's OS/VS2 (precursor to MVS, great-grandfather of zOS) which ran of the IBM S/370 mainframe. But the story doesn't stop there. DSS came from RSS which was present in some embryonic predecessor to VS2 back in the 1960s.

The problem with debugging technologies is that they have always been, and still are to some extent, considered to be poor relations to "the real operating system features". Debugging has been regarded by many and having no rightful place in a production system. This view as a general statement is nonsense and Sun's DTRACE has greatly helped dispel that nonsense view. Once could say that it has done Linux a great favour by bringing the need to a broader public arena. The work done by Sun has added to the debate and the work already done IBM, SGI, HP, Intel, Red Hat, SuSE and many others in this arena on Linux. The infrastructure to support dynamic instrumentation was accepted into the Linux kernel during the 2.5 development cycle. We now have comparable tool to DTRACE - namely System Tap, that exploits this infrastructure.

Richard J Moore - IBM Linux Technology Center


(Log in to post comments)

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 15, 2006 12:45 UTC (Fri) by zdzichu (subscriber, #17118) [Link]

Sorry, but SystemTap still isn't as stable as DTrace and can't trace userland. The amount of work which went in KProbes is, indeed, admirable.
But Sun after few years of work have production ready, useful product. And Linux people have not. Get that into your head.
Similar thing with ZFS. Solaris have it production ready, five years after first commit. Will we have something comparable in 2011? Doubtful.

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 15, 2006 15:20 UTC (Fri) by anonymous21 (guest, #30106) [Link]

Who cares about ZFS anyway?

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 24, 2006 21:59 UTC (Sun) by jamesd_wi (guest, #40720) [Link]

Who cares about ZFS anyway?

anyone who has ever used a LVM solution, anyone that knows in the future they may need to exceed the storage of a 64bit filesystem. Anyone that might want to take instaneous snapshots of there data.

It has veritas scared enough that it nows gives away its commercial filesystem that used to cost over $2000.

FreeBSD users are looking forward to having ZFS as well. (and are rapidly porting it.)

Even OSX users are thinking about getting ZFS.

Seems the only people that might claim that they don't care about ZFS is a linux zellot that tries to claim they don't want ZFS, but really they would love to have it but relize that their silly GPL license keep them from benefinting from ZFS's technology.

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 15, 2006 18:17 UTC (Fri) by giraffedata (subscriber, #1954) [Link]

The problem with debugging technologies is that they have always been, and still are to some extent, considered to be poor relations to "the real operating system features".

I think this stems from another concept: that it's wrong for production systems to have bugs. If you think by the time a system goes into production, it shouldn't have bugs, then diverting investment to production debugging tools feels wrong. It's a lot like distributing clean needles to illegal drug users and condoms in school.

I reached a conclusion a long time ago that this view is wrong. Bugs in the field should be expected and investment should be diverted from preventing bugs to making them easier to work around, diagnose, and repair in the field.

I'm seeing more and more diagnostics enabled by default, even with a runtime cost (memory, CPU time, disk space), indicating people are coming around to this view.

KHB: Dynamic Instrumentation of Production Systems (a.k.a. DTrace)

Posted Sep 17, 2006 15:18 UTC (Sun) by kreutzm (guest, #4700) [Link]

I agree to this view. When you run an evaluation of an IT product, it is quite common to find bugs (even in code shipping for some time already). You can either be proud about the bugs you've found or you can optimize your tools and get those bugs fixed quickly.


Copyright © 2017, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds