Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] [OMPI svn] svn:open-mpi r31376 - trunk/orte/mca/oob/tcp
From: Ralph Castain (rhc_at_[hidden])
Date: 2014-04-13 19:26:59


??? How is that any different/better than what was done?

On Apr 13, 2014, at 4:23 PM, George Bosilca <bosilca_at_[hidden]> wrote:

> Isn’t a socket or a pipe the right way to address this issue?
>
> George.
>
> On Apr 12, 2014, at 11:01 , svn-commit-mailer_at_[hidden] wrote:
>
>> Author: rhc (Ralph Castain)
>> Date: 2014-04-12 11:01:24 EDT (Sat, 12 Apr 2014)
>> New Revision: 31376
>> URL: https://svn.open-mpi.org/trac/ompi/changeset/31376
>>
>> Log:
>> Ensure we properly terminate the listening thread prior to exiting, but do so in a way that doesn't make us wait for select to timeout.
>>
>> Refs #4510
>>
>> Text files modified:
>> trunk/orte/mca/oob/tcp/oob_tcp_component.c | 9 ++++++---
>> trunk/orte/mca/oob/tcp/oob_tcp_component.h | 2 ++
>> trunk/orte/mca/oob/tcp/oob_tcp_listener.c | 17 ++++++++++++-----
>> 3 files changed, 20 insertions(+), 8 deletions(-)
>>
>> Modified: trunk/orte/mca/oob/tcp/oob_tcp_component.c
>> ==============================================================================
>> --- trunk/orte/mca/oob/tcp/oob_tcp_component.c Sat Apr 12 10:43:20 2014 (r31375)
>> +++ trunk/orte/mca/oob/tcp/oob_tcp_component.c 2014-04-12 11:01:24 EDT (Sat, 12 Apr 2014) (r31376)
>> @@ -140,8 +140,8 @@
>> if (ORTE_PROC_IS_HNP) {
>> OBJ_CONSTRUCT(&mca_oob_tcp_component.listen_thread, opal_thread_t);
>> mca_oob_tcp_component.listen_thread_active = false;
>> - mca_oob_tcp_component.listen_thread_tv.tv_sec = 0;
>> - mca_oob_tcp_component.listen_thread_tv.tv_usec = 300000;
>> + mca_oob_tcp_component.listen_thread_tv.tv_sec = 3600;
>> + mca_oob_tcp_component.listen_thread_tv.tv_usec = 0;
>> }
>> mca_oob_tcp_component.addr_count = 0;
>> OBJ_CONSTRUCT(&mca_oob_tcp_component.modules, opal_pointer_array_t);
>> @@ -659,7 +659,10 @@
>> ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
>>
>> if (ORTE_PROC_IS_HNP && mca_oob_tcp_component.listen_thread_active) {
>> - mca_oob_tcp_component.listen_thread_active = 0;
>> + mca_oob_tcp_component.listen_thread_active = false;
>> + /* tell the thread to exit */
>> + write(mca_oob_tcp_component.stop_thread, &i, sizeof(int));
>> + opal_thread_join(&mca_oob_tcp_component.listen_thread, NULL);
>> }
>>
>> while (NULL != (item = opal_list_remove_first(&mca_oob_tcp_component.listeners))) {
>>
>> Modified: trunk/orte/mca/oob/tcp/oob_tcp_component.h
>> ==============================================================================
>> --- trunk/orte/mca/oob/tcp/oob_tcp_component.h Sat Apr 12 10:43:20 2014 (r31375)
>> +++ trunk/orte/mca/oob/tcp/oob_tcp_component.h 2014-04-12 11:01:24 EDT (Sat, 12 Apr 2014) (r31376)
>> @@ -12,6 +12,7 @@
>> * Copyright (c) 2006-2013 Los Alamos National Security, LLC.
>> * All rights reserved.
>> * Copyright (c) 2010-2011 Cisco Systems, Inc. All rights reserved.
>> + * Copyright (c) 2014 Intel, Inc. All rights reserved
>> * $COPYRIGHT$
>> *
>> * Additional copyrights may follow
>> @@ -77,6 +78,7 @@
>> opal_thread_t listen_thread; /**< handle to the listening thread */
>> bool listen_thread_active;
>> struct timeval listen_thread_tv; /**< Timeout when using listen thread */
>> + int stop_thread; /**< file descriptor used to exit the listen thread */
>>
>> /* peers available via this transport - the index is the process name,
>> * and the pointer returned is the pointer to the last module that
>>
>> Modified: trunk/orte/mca/oob/tcp/oob_tcp_listener.c
>> ==============================================================================
>> --- trunk/orte/mca/oob/tcp/oob_tcp_listener.c Sat Apr 12 10:43:20 2014 (r31375)
>> +++ trunk/orte/mca/oob/tcp/oob_tcp_listener.c 2014-04-12 11:01:24 EDT (Sat, 12 Apr 2014) (r31376)
>> @@ -121,11 +121,11 @@
>> }
>> #endif
>>
>> - /* if I am the HNP, create a separate event base for the
>> - * listening thread so we can harvest connection requests
>> - * as rapidly as possible
>> + /* if I am the HNP, start a listening thread so we can
>> + * harvest connection requests as rapidly as possible
>> */
>> if (ORTE_PROC_IS_HNP) {
>> + mca_oob_tcp_component.stop_thread = open("/dev/null", O_RDWR);
>> mca_oob_tcp_component.listen_thread_active = true;
>> mca_oob_tcp_component.listen_thread.t_run = listen_thread;
>> mca_oob_tcp_component.listen_thread.t_arg = NULL;
>> @@ -641,15 +641,22 @@
>> FD_SET(listener->sd, &readfds);
>> max = (listener->sd > max) ? listener->sd : max;
>> }
>> + /* add the stop_thread fd */
>> + FD_SET(mca_oob_tcp_component.stop_thread, &readfds);
>> + max = (mca_oob_tcp_component.stop_thread > max) ? mca_oob_tcp_component.stop_thread : max;
>> +
>> /* set timeout interval */
>> timeout.tv_sec = mca_oob_tcp_component.listen_thread_tv.tv_sec;
>> timeout.tv_usec = mca_oob_tcp_component.listen_thread_tv.tv_usec;
>>
>> - /* Block in a select for a short (10ms) amount of time to
>> - * avoid hammering the cpu. If a connection
>> + /* Block in a select to avoid hammering the cpu. If a connection
>> * comes in, we'll get woken up right away.
>> */
>> rc = select(max + 1, &readfds, NULL, NULL, &timeout);
>> + if (!mca_oob_tcp_component.listen_thread_active) {
>> + /* we've been asked to terminate */
>> + return NULL;
>> + }
>> if (rc < 0) {
>> if (EAGAIN != opal_socket_errno && EINTR != opal_socket_errno) {
>> perror("select");
>> _______________________________________________
>> svn mailing list
>> svn_at_[hidden]
>> http://www.open-mpi.org/mailman/listinfo.cgi/svn
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
> Link to this post: http://www.open-mpi.org/community/lists/devel/2014/04/14500.php