Performance Monitoring Hardware

Hi All,

I was interested in doing some profiling to figure out basic block
execution frequencies and branch biases. This is the typical
information that one would need to do several profile based
optimizations (PBO).

An instrumentation based approach instruments the software to collect
profiling information. However, I am interested in using the existing
Performance Monitoring Hardware to do this. I am interested in finding
the frequencies of basic-block execution and their branch biases. The
reason why I dislike software based instrumentation and profiling is:
* It could get slow.
* It could require modifying the binary.
* It could probably perturb other temporal behvaiour of the program.

Using hardware based profiling is just more "transparent".

Right now I am not completely sure how to do this using the existing
perf. monitoring hardware in Pentium or the Itanium systems. AFAIK it
seems like existing h/w does counting of events and I am not sure how
I can get for example basic block frequencies?

Question:
Could someone tell me how performance monitoring hardware can be used
for basic-block profiling?

Eventually I want to feedback this profile information of BB profiles
and branch biases to the compiler (I am thinking gcc) to do static
PBO. Does anyone know any existing compiler infrastructure which does
static PBO using performance monitoring hardware for profiling?

Thanks!
Learner,

In comp.arch Learner wrote:
Eventually I want to feedback this profile information of BB profiles
and branch biases to the compiler (I am thinking gcc) to do static
PBO. Does anyone know any existing compiler infrastructure which does
static PBO using performance monitoring hardware for profiling?

It is possible that the Caliper tool for HP-UX/IPF does this for PBO
in the HP compilers. I cannot recall precisely if it is using the
Itanium PMU for the PBO part or not though, hence the possible.

Some details might be found at http://www.hp.com/go/caliper

rick jones
--
Wisdom Teeth are impacted, people are affected by the effects of events.
these opinions are mine, all mine; HP might not want them anyway...
feel free to post, OR email to raj in cup.hp.com but NOT BOTH...

Learner wrote:
I was interested in doing some profiling to figure out basic block
execution frequencies and branch biases. This is the typical
information that one would need to do several profile based
optimizations (PBO).

An instrumentation based approach instruments the software to collect
profiling information. However, I am interested in using the existing
Performance Monitoring Hardware to do this. I am interested in finding
the frequencies of basic-block execution and their branch biases. The
reason why I dislike software based instrumentation and profiling is:
* It could get slow.
* It could require modifying the binary.
* It could probably perturb other temporal behvaiour of the program.

Using hardware based profiling is just more "transparent".

Right now I am not completely sure how to do this using the existing
perf. monitoring hardware in Pentium or the Itanium systems. AFAIK it
seems like existing h/w does counting of events and I am not sure how
I can get for example basic block frequencies?

On Itanium/Itanium2, setup the BTB to count all branches (or just
the taken ones if you have enough logic to infer the not-taken ones from
the taken), program BRANCH_EVENTS to overflow in one of the counting
PMD/PMC (4-7), and off you go.

That's pretty much what 'caliper branch_prediction ...' does.
URL:http://www.hp.com/go/hpcaliper
(currently only on HP-UX although ..., nah, I can't say that)

Note however that the PBO data generated by Caliper is collected
via 'precise' instrumentation (although dynamic instrumentation
of more-or-less unprepared binaries), although it can be completed
by some PMU data.

Question:
Could someone tell me how performance monitoring hardware can be used
for basic-block profiling?

See above.

On Linux Itanium, check Stephane Eranian's perfmon:
URL:http://www.hpl.hp.com/research/linux/perfmon/pfmon.php4

If you want something similar on HP-UX Itanium, feel free to contact me.
A similar API exists but is currently left undocumented as it is
still rapidly evolving.

Eventually I want to feedback this profile information of BB profiles
and branch biases to the compiler (I am thinking gcc) to do static
PBO. Does anyone know any existing compiler infrastructure which does
static PBO using performance monitoring hardware for profiling?

No comment.

Eric

(Learner) wrote in message . com...
Hi All,

I was interested in doing some profiling to figure out basic block
execution frequencies and branch biases. This is the typical
information that one would need to do several profile based
optimizations (PBO).

An instrumentation based approach instruments the software to collect
profiling information. However, I am interested in using the existing
Performance Monitoring Hardware to do this. I am interested in finding
the frequencies of basic-block execution and their branch biases. The
reason why I dislike software based instrumentation and profiling is:
* It could get slow.
* It could require modifying the binary.
* It could probably perturb other temporal behvaiour of the program.

What do you mean by perturb? You think that PGO doesn't take into
account the impact that the monitoring has on instruction count etc.?

Using hardware based profiling is just more "transparent".

Right now I am not completely sure how to do this using the existing
perf. monitoring hardware in Pentium or the Itanium systems. AFAIK it
seems like existing h/w does counting of events and I am not sure how
I can get for example basic block frequencies?

One word: VTune (http://www.intel.com/software/products/vtune/index.htm)

Now a note about VTune. You cannot count ALL branches, without being
just as slow as PGO. VTune uses sampling, which is fairly effective
(from what I hear and read) at getting accurate figures, but it isn't
as accurate as counting all branches. It is also much faster, perhaps
having overhead costs of 5%.

You should be able to count the number of branches and the number of
mispredicts, but I don't know if you can count taken vs. not taken.

Question:
Could someone tell me how performance monitoring hardware can be used
for basic-block profiling?

VTune has a feature called Call Graph Profiling which should do almost
exactly what you want. It's built into the application and is pretty
easy to do, esp. if you have the source code.

David