Remote update of values

[Mert Hidayetoglu]

 

Hi all,

 

I am trying to understand the AM mechanism for updating remote values. I wrote a simple program below to block remote process until the AM function is called on the receiver side. However, the receiver process never gets out of the while loop.

 

Is there a way to make sure to provide consistency without using a barrier?

 

int complete = 1;

void am_handler (gex_Token_t token) { complete  = 1; }

 

int main() {

  …

  int sender = 3;

  int recver = 0;

  if (myrank == recver)

   complete = 0;

  gex_Event_Wait(gex_Coll_barrierNB(myteam, 0));

 

  if (myrank == sender)

    gex_AM_RequestShort0 (myteam, recver, GEX_AM_INDEX_BASE, 0);

  while (!complete);

 

  gex_Event_Wait (gex_Coll_barrierNB(myteam, 0));

}

 

I appreciate any help to understand proper way of updating remote values.

 

Best regards,

Mert

[Mert Hidayetoglu]

I understand that I need to use HSL for ensuring consistency: https://gasnet.lbl.gov/docs/gasnet.html#Handler_002dSafe-Locks

 

I updated my code accordingly, but I am still missing something because the receiver process does not break out of the loop.

 

Am I placing the locks & unlocks at the right locations?

 

int complete = 1;

gex_HSL_t hsl = GEX_HSL_INITIALIZER;

void am_handler (gex_Token_t token) {

  gex_HSL_Lock(&hsl);

  complete  = 1;

  gex_HSL_Unlock(&hsl);

}

 

int main() {

  …

  int sender = 3;

  int recver = 0;

  if (myrank == recver)

    complete = 0;

  gex_Event_Wait(gex_Coll_barrierNB(myteam, 0));

 

  if (myrank == sender)

    gex_AM_RequestShort0 (myteam, recver, GEX_AM_INDEX_BASE, 0);

 

  do {

    gex_HSL_Lock(&hsl);

    if (complete) {

      gex_HSL_Unlock(&hsl); // leave no HSL locked

      break;

    }

    gex_HSL_Unlock(&hsl);

  } while (true)

 

  gex_Event_Wait (gex_Coll_barrierNB(myteam, 0));

}

 

I appreciate any help to understand proper way of using HSL mechanism.

 

Best regards,

Mert

[Paul H. Hargrove]

Mert,

Active Messages do not, in general, run asynchronously.

You need to call `gasnet_AMPoll()` or a blocking function (such as gex_Event_Wait()) to process in-bound AM Requests and Replies.

The optional receive progress thread in ibv-conduit is an exception, not present for other networks.

You are correct about the need for HSLs if you were doing a non-trivial state change, in particular if multiple threads are running.

However, I believe you can safely eliminate the HSL for the simple case of a `complete` transitioning from 0 to 1.

You could write

    while (!complete) gasnet_AMPoll();

or

   GASNET_BLOCKUNTIL(complete);

If there is a non-trivial transition

   GASNET_BLOCKUNTIL(is_complete());

Where is_complete() is

  int is_complete(void) {

     gex_HSL_Lock(&hsl);

     int result = complete;

     gex_HSL_Unlock(&hsl);

     return result;

  }

I hope that helps.

Have a good weekend,

-Paul

--
You received this message because you are subscribed to the Google Groups "gasnet-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gasnet-users...@lbl.gov.
To view this discussion on the web visit https://groups.google.com/a/lbl.gov/d/msgid/gasnet-users/SJ0PR02MB73575C9F36D9A2C4E6D41CF9D1FC2%40SJ0PR02MB7357.namprd02.prod.outlook.com.

--

Paul H. Hargrove <PHHar...@lbl.gov>
Pronouns: he, him, his

Computer Languages & Systems Software (CLaSS) Group

Computer Science Department

Lawrence Berkeley National Laboratory

[Mert Hidayetoglu]

Thanks for your help, Paul! gasnet_AMPoll() solves the problem. As you said, the update in my example is trivial, so no need for HSL unless multiple threads try to update the same value.

 

“

You could write

    while (!complete) gasnet_AMPoll();

or

   GASNET_BLOCKUNTIL(complete);

”

 

I found GASNET_BLOCKUNTIL macro handy. What are the implications of choosing one over the other? Should I assume the macro is preferred over looping over gasnet_AMPoll() in this case?

 

Mert

[Paul H. Hargrove]

Mert,

I recommend the GASNET_BLOCKUNTIL macro.

In the future, implementations on some systems may be able to block rather than spin poll, reducing energy usage and/or resource contention.

-Paul

[Mert Hidayetoglu]

Noted with thanks!

Mert

---

From: Mert Hidayetoglu <me...@stanford.edu>
Sent: Saturday, June 1, 2024 2:25:02 AM
To: Paul H. Hargrove <phhar...@lbl.gov>