From: Michael Z. <zmi...@ri...> - 2002-09-12 13:03:10
|
On Thursday 12 September 2002 15:44, John Levon wrote: > On Thu, Sep 12, 2002 at 11:46:09AM +0300, Michael Zayats wrote: >=20 > > > But I'm a bit surprised > > > you're considering PEBS support already, seeing as that will requir= e > > > more significant oprofile changes than the periodic interrupt style= =2E > >=20 > > do you mean RTC?=20 >=20 > No, I mean the perfctr-based interrupt approach similar to P6 that we > have already. >=20 > Basically there are two modes in the Pentium 4 - the interrupt mode, > which is very similar to the current system - you set up the counters > and they send an interrupt on overflow, and we read the EIP and store i= t > ourselves. IEBS other words, this can be ported to, but if we have PEBS we don't nee= d=20 IEBS >=20 > PEBS is different because you create a buffer and set up the performanc= e > counter stuff to "handle itself" to some degree.=20 AFAIK, it's writing to some buffer instead giving an interrupt and rises= an=20 interrupt when buffer overflows. Since the whole PEBS management already=20 written in ABYSS, I don't expect it to be of any major problem. > This is less like > oprofile already is, so would probably require more work. >=20 > > My biggest concern there is the NMI related issues, can somebody tell= me=20 more=20 > > about specifics of NMI, and what problems he expects to arise for P4?= =20 >=20 > There shouldn't be any NMI-specific problems for the PEntium 4, at leas= t > not for the non-PEBS approach. I forget if you set up NMI delivery in > the same way (on P6, you write to LVTPC on the local APIC to tell it to > use NMI delivery). basically if you forget setting it NMI, it will not profile whenever inte= rrupt=20 are blocked, right? >=20 > The major work should really be describing the counter setup and > initialisation, which is quite a bit different. >=20 As for now I will try to get one counter working, submit a patch and only= then=20 will work on the rest of the framework.=20 > regards > john > --=20 > "This *is* Usenet, after all, where virtually every conversation that g= oes=20 on > is fairly ludicrous in the first place." > =09- Godwin's Law FAQ >=20 >=20 --=20 Michael Zayats Riverhead ltd. |