Why systemtap loses events and what can be done about it?

[Frank Ch. Eigler]

Hi -

On Tue, Sep 04, 2007 at 11:53:58AM -0700, Brad Peters wrote:

> [...]  Definitely an interesting idea; Given Frank's concern, it
> seems that all data collection would have to take place in the
> front-end handler.

Yeah, but that cancels one of the possible advantages of deferred
handling: the ability to sleep within probe handlers.  Some data
requires sleeping to extract "reliably" (in terms of data correctness;
unfortunately this tends to introduce deadlocks, so we don't do it).

> The Bluedye team (Mike Strosaker, Nathan Fontenot and myself)
> implemented something similar, collecting data immediately after a
> probe is hit, then using a work queue to do all relay channel writes
> at a later time (i.e., when interrupts are enabled).

(A quick google search can't find any public info about BlueDye.)

> I have no idea what time-savings we saw, but I will find some time
> today to hack out our work-queue code and look at the timing impact

Thanks, should be interesting.

- FChE