The 3.16 kernel has been released

Posted Aug 7, 2014 2:20 UTC (Thu) by Cyberax (✭ supporter ✭, #52523)
In reply to: The 3.16 kernel has been released by kloczek
Parent article: The 3.16 kernel has been released

> Do you really think that reporting CPC registers data it is the same what you can do in few lines D script correlating CPC data with few other things?
SystemTap can do this just fine: https://github.com/fche/systemtap/blob/master/testsuite/s...

> Please don't take this personally but seems you are yet another person which does not fully understand technological impact of approach implemented in DTrace.
I'd used DTrace. It's nice but not groundshaking. And even before perf on Linux, I used oprofile and other similar tools to find bottlenecks in my apps.

The 3.16 kernel has been released

Posted Aug 7, 2014 3:06 UTC (Thu) by kloczek (guest, #6391) [Link] (4 responses)

> SystemTap can do this just fine: https://github.com/fche/systemtap/blob/master/testsuite/s...

Did you watch Big Bang Theory when Sheldon explained Penny what it means "just fine"? :)

https://www.youtube.com/watch?v=Yo-CWXQ8_1M
https://www.youtube.com/watch?v=_amwWlgS6LM

Try to imagine that my reaction on "just fine" phrase is like Penny reaction :P

The 3.16 kernel has been released

Posted Aug 7, 2014 3:16 UTC (Thu) by Cyberax (✭ supporter ✭, #52523) [Link] (3 responses)

No, I don't watch BBT and I don't really want to. So please explain with examples why DTrace is so much superior in this case.

The 3.16 kernel has been released

Posted Aug 7, 2014 4:15 UTC (Thu) by kloczek (guest, #6391) [Link] (2 responses)

> So please explain with examples why DTrace is so much superior in this case

Because you can do base processing huge volume of tracing data in place of hook using D code instead doing this offline.
Shuffling big volumes of data from kernel space to user space is causing kind of observability quantum effect (observer object state is disturbed by observation).
DTrace it is not like perf which is event driven.
Perf is provides analysis tools to navigate in large multi-GB traces. Dtrace does not have this because is designed to use concise traces. Simple perf is more about offline than online analysis.
So far systemtap or perf does not provides users space providers.
DTrace on Linux now is able to use USDT providers.
Example: https://blogs.oracle.com/wim/entry/mysql_5_6_20_4

The 3.16 kernel has been released

Posted Aug 7, 2014 4:18 UTC (Thu) by Cyberax (✭ supporter ✭, #52523) [Link] (1 responses)

Uhm, no. SystemTap compiles probes into kernel modules, so there's no userspace-kernelspace overhead. And perf subsystem also supports filters.

The 3.16 kernel has been released

Posted Aug 9, 2014 12:13 UTC (Sat) by kloczek (guest, #6391) [Link]

> SystemTap compiles probes into kernel modules, so there's no userspace-kernelspace overhead.

How many times did you rash system using SystemTap?
I had such situations hundreds times.
This is why DTrace VM doing whole work is better.

> And perf subsystem also supports filters.

You don't understand where is the problem.
If instrumentation generates even this event must be queued. Effectively here you will have few context switches. Taking data from queue add more cs and take event from queue to discard part events will be always wast of CPU/time.

Now PCI bus cannot handle more than about 300k ops/s but new PCI protocol may change this to millions/s. Try to imagine how big overhead may be after this compare to DTrace way on for example tracing IOs.