Charm++ parallel programming system

[Ronak Buch <rabuch2 AT illinois.edu>]

Yes, you can basically ignore the rest of the fields (they're some mixture of ordering, IDs, performance counters, etc.). The trailing zeros are some of these fields and can usually be ignored.

dummy_thread_ep is basically a catchall for things that are traced that the runtime doesn't have enough information for to attribute to something else. It can be meaningful, but it could be a lot of different things depending on what exactly your program is doing, so it's hard for me to make any conclusions without more information. Usually looking at the bookending events that have meaningful attributions is enough to figure out what's going on in those sections.

On Tue, Jun 5, 2018 at 5:51 PM, Rob Van der Wijngaart <robv AT nvidia.com> wrote:

Thanks much, Ronak!

Below is a representative snippet from one of my log files.

The lines that start with 2 or 3 have 8 and 13 fields each. I guess I can ignore the fields that you’re not describing. A lot of the records that start with 2 or 3 refer to ENTRY CHARE 0, which refers to dummy_thread_ep, according to the sts file (the line says: “ENTRY CHARE 0 dummy_thread_ep 0 0”). Is that a meaningful mapping, and do the trailing zeroes have any significance? Thanks.

 

Rob

 

3 0 355 14898768 569 16 0 11638933

2 4 0 14898768 2950 14 0 0 14 369 0 0 0

3 0 0 14898864 2950 14 0 11638933

14 14898866 14

15 14901654 14

2 5 355 14901657 893 16 100 0 0 0 0 0 11641931

1 4 0 14901667 2951 14 0 0

3 0 355 14901669 893 16 0 11641931

2 5 355 14901669 900 16 100 0 0 0 0 0 11641931

1 4 0 14901679 2952 14 0 0

3 0 355 14901681 900 16 0 11641931

2 4 0 14901681 2952 14 0 0 14 2118 0 0 0

3 0 0 14901772 2952 14 0 11641931

2 5 355 14901775 954 16 100 0 0 0 0 0 11641931

1 4 0 14901787 2953 14 0 0

3 0 355 14901787 954 16 0 11641931

2 4 0 14901789 2953 14 0 0 14 1958 0 0 0

3 0 0 14901878 2953 14 0 11641931

2 5 355 14901880 1019 16 100 0 0 0 0 0 11641931

1 4 0 14901890 2954 14 0 0

3 0 355 14901894 1019 16 0 11641931

2 5 355 14901895 1006 16 100 0 0 0 0 0 11641931

1 4 0 14901906 2955 14 0 0

3 0 355 14901906 1006 16 0 11641931

2 4 0 14901908 2955 14 0 0 14 1653 0 0 0

3 0 0 14901993 2955 14 0 11641931

2 4 0 14901995 2955 14 0 0 14 1349 0 0 0

3 0 0 14902085 2955 14 0 11641931

2 4 0 14902086 2955 14 0 0 14 1796 0 0 0

3 0 0 14902175 2955 14 0 11641931

14 14902176 14

15 14902230 14

 

 

From: Ronak Buch <rabuch2 AT illinois.edu>
Sent: Tuesday, June 5, 2018 3:09 PM

To: Rob Van der Wijngaart <robv AT nvidia.com>
Cc: charm AT cs.illinois.edu
Subject: Re: [charm] Obtaining Charm++ profile

 

No, we don't have a good description of the record format right now other than the source code that generates it in Charm (src/ck-perf/trace-projections.C) and the source code that parses it in Projections (GenericLogReader.java). I'll provide a brief summary below, but if there's something specific that you want, I can explain it.

 

In general, each line in the Projections log file represents a specific event. The ones that you're probably interested in are those beginning with 2 or 3, which represent BEGIN_PROCESSING and END_PROCESSING events (the start and end of entry methods). The third field in those lines represents the entry ID, which can be converted a function name by using the sts file (in which ENTRY CHARE lines provide a mapping between an entry ID and the actual name of the entry method). The fourth field represents the timestamp of the event (so the difference between the two values in this position for corresponding BEGIN_PROCESSING and END_PROCESSING events is the total time of the entry method). The sixth value is the PE, which indicates which PE served as that event's source. The seventh field for BEGIN_PROCESSING events is the message length.

 

Hopefully that helps.

 

Thanks,

Ronak

 

On Tue, Jun 5, 2018 at 4:38 PM, Rob Van der Wijngaart <robv AT nvidia.com> wrote:

OK, I got my .log, .sts, and .projrc files for my run! Next question, if you don’t mind. I know of Projections, but I’d like to use my own graph analysis tools to parse the log files. Is there a description of the records in the .log and .sts files? Thanks!

 

Rob

 

From: Rob Van der Wijngaart
Sent: Tuesday, June 5, 2018 2:11 PM
To: 'Ronak Buch' <rabuch2 AT illinois.edu>
Cc: charm AT cs.illinois.edu
Subject: RE: [charm] Obtaining Charm++ profile

 

Thanks, Ronak!

 

Rob

 

From: Ronak Buch <rabuch2 AT illinois.edu>
Sent: Tuesday, June 5, 2018 2:05 PM
To: Rob Van der Wijngaart <robv AT nvidia.com>
Cc: charm AT cs.illinois.edu
Subject: Re: [charm] Obtaining Charm++ profile

 

You built Charm++ with the correct flag ("--enable-tracing"). The issue is that the application, when linked, needs the "-tracemode projections" flag.

 

Specifically, if you're building NAMD (as you seem to be), you can just run "make projections," which includes the "-tracemode projections" flag already.

 

Thanks,

Ronak

 

 

On Tue, Jun 5, 2018 at 12:47 PM, Rob Van der Wijngaart <robv AT nvidia.com> wrote:

Hello Team,

 

I am trying to obtain a profile of a NAMD run using Charm++. I built Charm++ this way:

./build charm++ multicore-linux64 gcc --with-production --enable-tracing -j40 -tracemode summary

According to the man page I should now automatically obtain a profile if I run the application like this:

./namd-2.13-multicore-cuda +traceroot $PWD +p32 +setcpuaffinity +pemap 0-31 +devices 0,1,2,3 stmv/stmv.namd

However, I do not see any profile. If I then add +sumonly to the command line, I get this error message:

WARNING: +sumonly is a command line argument beginning with a '+' but was not parsed by the RTS.

If any of the above arguments were intended for the RTS you may need to recompile Charm++ with different options.

So it appears I did not build Charm++ properly to support tracing. Can you tell me how to do it right?

Second, I am actually interesting in doing projections, not obtaining a summary profile, but I do not understand this instruction in the manual:

Link time option: -tracemode projections

After all, I only issue a build command to create the runtime, with no explicit link calls issued. Can you advise?
This concerns v6.8.2, which I obtained as a tar ball.

Thanks,
Rob

 

 

---

This email message is for the sole use of the intended recipient(s) and may contain confidential information.  Any unauthorized review, use, disclosure or distribution is prohibited.  If you are not the intended recipient, please contact the sender by reply email and destroy all copies of the original message.

---