ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
9cfc408fd4
openmpi/ompi/mca/btl/usnic/btl_usnic_mca.c

291 строка
12 KiB
C
Исходник Обычный вид История

First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
/*
* Copyright (c) 2004-2008 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2011 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006 Sandia National Laboratories. All rights
* reserved.
* Copyright (c) 2008-2013 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2012 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#ifdef HAVE_STRING_H
#include <string.h>
#endif
#include <errno.h>
#include <infiniband/verbs.h>
#include "opal/mca/base/mca_base_var.h"
#include "opal/util/argv.h"
#include "ompi/constants.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/btl/base/base.h"
#include "ompi/mca/common/verbs/common_verbs.h"
#include "btl_usnic.h"
#include "btl_usnic_frag.h"
#include "btl_usnic_endpoint.h"
#include "btl_usnic_module.h"
/*
* Local flags
*/
enum {
REGINT_NEG_ONE_OK = 0x01,
REGINT_GE_ZERO = 0x02,
REGINT_GE_ONE = 0x04,
REGINT_NONZERO = 0x08,
REGINT_MAX = 0x88
};
enum {
REGSTR_EMPTY_OK = 0x01,
REGSTR_MAX = 0x88
};
/*
* utility routine for string parameter registration
*/
static int reg_string(const char* param_name,
const char* help_string,
const char* default_value, char **storage,
int flags, int level)
{
*storage = (char*) default_value;
mca_base_component_var_register(&mca_btl_usnic_component.super.btl_version,
param_name, help_string,
MCA_BASE_VAR_TYPE_STRING,
NULL,
0,
0,
level,
MCA_BASE_VAR_SCOPE_READONLY,
storage);
if (0 == (flags & REGSTR_EMPTY_OK) &&
(NULL == *storage || 0 == strlen(*storage))) {
opal_output(0, "Bad parameter value for parameter \"%s\"",
param_name);
return OMPI_ERR_BAD_PARAM;
}
return OMPI_SUCCESS;
}
/*
* utility routine for integer parameter registration
*/
static int reg_int(const char* param_name,
const char* help_string,
int default_value, int *storage, int flags, int level)
{
*storage = default_value;
mca_base_component_var_register(&mca_btl_usnic_component.super.btl_version,
param_name, help_string,
MCA_BASE_VAR_TYPE_INT,
NULL,
0,
0,
level,
MCA_BASE_VAR_SCOPE_READONLY,
storage);
if (0 != (flags & REGINT_NEG_ONE_OK) && -1 == *storage) {
return OMPI_SUCCESS;
}
if ((0 != (flags & REGINT_GE_ZERO) && *storage < 0) ||
(0 != (flags & REGINT_GE_ONE) && *storage < 1) ||
(0 != (flags & REGINT_NONZERO) && 0 == *storage)) {
opal_output(0, "Bad parameter value for parameter \"%s\"",
param_name);
return OMPI_ERR_BAD_PARAM;
}
return OMPI_SUCCESS;
}
Add usnic connectivity-checking agent service. Basically: since usnic is a connectionless transport, we do not get OS-provided services "for free" that connection-oriented transports get, namely: "hey, I wasn't able to make a connection to peer X", and "hey, your connection to peer X has died." This connectivity-checker runs in a separate progress thread in the usnic BTL in local rank 0 on each server. Upon first send in any process, the connectivty-checker agent will send some UDP pings to the peer to ensure that we can reach it. If we can't, we'll abort the job with a nice show_help message. There's a lengthy comment in btl_usnic_connectivity.h explains the scheme and how it works. Reviewed by Dave Goodell. cmr=v1.7.5:ticket=trac:4253 This commit was SVN r30860. The following Trac tickets were found above: Ticket 4253 --> https://svn.open-mpi.org/trac/ompi/ticket/4253
2014-02-27 02:21:25 +04:00
/*
* utility routine for integer parameter registration
*/
static int reg_bool(const char* param_name,
const char* help_string,
bool default_value, bool *storage, int level)
{
*storage = default_value;
mca_base_component_var_register(&mca_btl_usnic_component.super.btl_version,
param_name, help_string,
MCA_BASE_VAR_TYPE_BOOL,
NULL,
0,
0,
level,
MCA_BASE_VAR_SCOPE_READONLY,
storage);
return OMPI_SUCCESS;
}
First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
int ompi_btl_usnic_component_register(void)
{
Minor improvements to the usnic BTL: 1. Fix ompi_info memory leak in usnic BTL: do not allocate memory in the component register function, because ompi_info only calls the component register function and then dlclose's the component -- it does not call component finalize. Instead, defer parsing the MCA param (and alloc'ing memory) until the component init function so that any allocated memory can be freed in the component close function. 1. Also add a new check to ensure that we actually have some part numbers to check. Add a show_help message if we don't find any vendor part IDs to check. 1. Add a verbose output if usnic disqualifies itself from selection because THREAD_MULTIPLE was specified. cmr=v1.7.5:reviewer=dgoodell This commit was SVN r30073.
2013-12-24 15:57:35 +04:00
int tmp, ret = 0;
First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
static int max_modules;
static int stats_relative;
static int want_numa_device_assignment;
static int sd_num;
static int rd_num;
static int prio_sd_num;
static int prio_rd_num;
static int cq_num;
static int max_tiny_payload;
static int eager_limit;
static int rndv_eager_limit;
usnic: pack via convertor on the fly If we need to use a convertor, go back to stashing that convertor in the frag and populating segments "on the fly" (in ompi_btl_usnic_module_progress_sends). Previously we would pack into a chain of chunk segments at prepare_src time, unnecessarily consuming additional memory. Reviewed-by: Jeff Squyres <jsquyres@cisco.com> Reviewed-by: Reese Faucette <rfaucett@cisco.com> This commit was SVN r29592.
2013-11-05 02:52:03 +04:00
static int pack_lazy_threshold;
First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
#define CHECK(expr) do {\
tmp = (expr); \
if (OMPI_SUCCESS != tmp) ret = tmp; \
} while (0)
CHECK(reg_int("max_btls",
"Maximum number of usNICs to use (default: 0 = as many as are available)",
0, &max_modules,
REGINT_GE_ZERO, OPAL_INFO_LVL_2));
mca_btl_usnic_component.max_modules = (size_t) max_modules;
CHECK(reg_string("if_include",
"Comma-delimited list of devices/networks to be used (e.g. \"usnic_0,10.10.0.0/16\"; empty value means to use all available usNICs). Mutually exclusive with btl_usnic_if_exclude.",
NULL, &mca_btl_usnic_component.if_include,
REGSTR_EMPTY_OK, OPAL_INFO_LVL_1));
CHECK(reg_string("if_exclude",
"Comma-delimited list of devices/networks to be excluded (empty value means to not exclude any usNICs). Mutually exclusive with btl_usnic_if_include.",
NULL, &mca_btl_usnic_component.if_exclude,
REGSTR_EMPTY_OK, OPAL_INFO_LVL_1));
CHECK(reg_int("stats",
"A non-negative integer specifying the frequency at which each USNIC BTL will output statistics (default: 0 seconds, meaning that statistics are disabled)",
0, &mca_btl_usnic_component.stats_frequency, 0,
OPAL_INFO_LVL_4));
mca_btl_usnic_component.stats_enabled =
(bool) (mca_btl_usnic_component.stats_frequency > 0);
CHECK(reg_int("stats_relative",
"If stats are enabled, output relative stats between the timestemps (vs. cumulative stats since the beginning of the job) (default: 0 -- i.e., absolute)",
0, &stats_relative, 0, OPAL_INFO_LVL_4));
mca_btl_usnic_component.stats_relative = (bool) stats_relative;
CHECK(reg_string("mpool", "Name of the memory pool to be used",
"grdma", &mca_btl_usnic_component.usnic_mpool_name, 0,
OPAL_INFO_LVL_5));
want_numa_device_assignment = OPAL_HAVE_HWLOC ? 1 : -1;
CHECK(reg_int("want_numa_device_assignment",
"If 1, use only Cisco VIC ports thare are a minimum NUMA distance from the MPI process for short messages. If 0, use all available Cisco VIC ports for short messages. This parameter is meaningless (and ignored) unless MPI proceses are bound to processor cores. Defaults to 1 if NUMA support is included in Open MPI; -1 otherwise.",
want_numa_device_assignment,
&want_numa_device_assignment,
0, OPAL_INFO_LVL_5));
mca_btl_usnic_component.want_numa_device_assignment =
(1 == want_numa_device_assignment) ? true : false;
CHECK(reg_int("sd_num", "Maximum send descriptors to post (-1 = pre-set defaults; depends on number and type of devices available)",
-1, &sd_num, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.sd_num = (int32_t) sd_num;
CHECK(reg_int("rd_num", "Number of pre-posted receive buffers (-1 = pre-set defaults; depends on number and type of devices available)",
-1, &rd_num, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.rd_num = (int32_t) rd_num;
CHECK(reg_int("prio_sd_num", "Maximum priority send descriptors to post (-1 = pre-set defaults; depends on number and type of devices available)",
-1, &prio_sd_num, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.prio_sd_num = (int32_t) prio_sd_num;
CHECK(reg_int("prio_rd_num", "Number of pre-posted priority receive buffers (-1 = pre-set defaults; depends on number and type of devices available)",
-1, &prio_rd_num, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.prio_rd_num = (int32_t) prio_rd_num;
CHECK(reg_int("cq_num", "Number of completion queue entries (-1 = pre-set defaults; depends on number and type of devices available; will error if (sd_num+rd_num)>cq_num)",
-1, &cq_num, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.cq_num = (int32_t) cq_num;
CHECK(reg_int("retrans_timeout", "Number of microseconds before retransmitting a frame",
1000, &mca_btl_usnic_component.retrans_timeout,
REGINT_GE_ONE, OPAL_INFO_LVL_5));
CHECK(reg_int("priority_limit", "Max size of \"priority\" messages (0 = use pre-set defaults; depends on number and type of devices available)",
0, &max_tiny_payload,
REGINT_GE_ZERO, OPAL_INFO_LVL_5));
ompi_btl_usnic_module_template.max_tiny_payload =
(size_t) max_tiny_payload;
CHECK(reg_int("eager_limit", "Eager send limit (0 = use pre-set defaults; depends on number and type of devices available)",
0, &eager_limit, REGINT_GE_ZERO, OPAL_INFO_LVL_5));
ompi_btl_usnic_module_template.super.btl_eager_limit = eager_limit;
CHECK(reg_int("rndv_eager_limit", "Eager rendezvous limit (0 = use pre-set defaults; depends on number and type of devices available)",
0, &rndv_eager_limit, REGINT_GE_ZERO, OPAL_INFO_LVL_5));
ompi_btl_usnic_module_template.super.btl_rndv_eager_limit =
rndv_eager_limit;
usnic: pack via convertor on the fly If we need to use a convertor, go back to stashing that convertor in the frag and populating segments "on the fly" (in ompi_btl_usnic_module_progress_sends). Previously we would pack into a chain of chunk segments at prepare_src time, unnecessarily consuming additional memory. Reviewed-by: Jeff Squyres <jsquyres@cisco.com> Reviewed-by: Reese Faucette <rfaucett@cisco.com> This commit was SVN r29592.
2013-11-05 02:52:03 +04:00
CHECK(reg_int("pack_lazy_threshold", "Convertor packing on-the-fly threshold (-1 = always pack eagerly, 0 = always pack lazily, otherwise will pack on the fly if fragment size is > limit)",
USNIC_DFLT_PACK_LAZY_THRESHOLD, &pack_lazy_threshold, REGINT_NEG_ONE_OK, OPAL_INFO_LVL_5));
mca_btl_usnic_component.pack_lazy_threshold = pack_lazy_threshold;
This is what r30852 should have been: Consolidate into a single, outter loop of ibv_create_ah() calls Follow on to SVN trunk r30850: consolidate the ibv_create_ah() calls into a single loop, MPI_WAITALL-style. That is, call the (effectively non-blocking) ibv_create_ah() for each endpoint. If we get NULL+EAGAIN, it means that the UDP ARP is still ongoing down in the kernel, so just try again later. We put these all into a single loop because it allows us to parallelize the ARP progress in the kernel. cmr=v1.7.5:ticket=trac:4253 This commit was SVN r30879. The following SVN revision numbers were found above: r30850 --> open-mpi/ompi@3641500442c1362e148d0d6272db91d75ec92116 r30852 --> open-mpi/ompi@4e282a329578c9e7296d5a3809fedfa9f83870df The following Trac tickets were found above: Ticket 4253 --> https://svn.open-mpi.org/trac/ompi/ticket/4253
2014-02-27 21:19:50 +04:00
CHECK(reg_int("arp_timeout", "Timeout, in seconds, for the maximum delay between ARP replies when using the usNIC/UDP transport (ignored when using the usNIC/L2 transport, must be >=1)",
10, &mca_btl_usnic_component.arp_timeout,
REGINT_GE_ONE, OPAL_INFO_LVL_6));
First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
/* Default to bandwidth auto-detection */
ompi_btl_usnic_module_template.super.btl_bandwidth = 0;
ompi_btl_usnic_module_template.super.btl_latency = 4;
Add usnic connectivity-checking agent service. Basically: since usnic is a connectionless transport, we do not get OS-provided services "for free" that connection-oriented transports get, namely: "hey, I wasn't able to make a connection to peer X", and "hey, your connection to peer X has died." This connectivity-checker runs in a separate progress thread in the usnic BTL in local rank 0 on each server. Upon first send in any process, the connectivty-checker agent will send some UDP pings to the peer to ensure that we can reach it. If we can't, we'll abort the job with a nice show_help message. There's a lengthy comment in btl_usnic_connectivity.h explains the scheme and how it works. Reviewed by Dave Goodell. cmr=v1.7.5:ticket=trac:4253 This commit was SVN r30860. The following Trac tickets were found above: Ticket 4253 --> https://svn.open-mpi.org/trac/ompi/ticket/4253
2014-02-27 02:21:25 +04:00
/* Connectivity verification */
mca_btl_usnic_component.connectivity_enabled = true;
CHECK(reg_bool("connectivity_check",
"Whether to enable the usNIC connectivity check upon first send (default = 1, enabled; 0 = disabled)",
mca_btl_usnic_component.connectivity_enabled,
&mca_btl_usnic_component.connectivity_enabled,
OPAL_INFO_LVL_3));
btl_usnic_mca.c: Increase default connectivity checker frequency In abusive MPI communication patterns, sending a UDP ping only once a second may not be sufficient -- all the UDP pings may be dropped. So increase the frequency of the pings to every quarter second, and allow more total pings to be sent. Total timeout time is still the same (10 seconds) -- we'll just now try 40 times (i.e., once every quarter second) as opposed to 10 times (i.e., once a second). Testing has shown that this frequency allows the connectivity checker to always succeed even in the many-to-one abusive communication patterns. cmr=v1.8.2:reviewer=dgoodell This commit was SVN r31602.
2014-05-02 15:06:18 +04:00
mca_btl_usnic_component.connectivity_ack_timeout = 250;
Add usnic connectivity-checking agent service. Basically: since usnic is a connectionless transport, we do not get OS-provided services "for free" that connection-oriented transports get, namely: "hey, I wasn't able to make a connection to peer X", and "hey, your connection to peer X has died." This connectivity-checker runs in a separate progress thread in the usnic BTL in local rank 0 on each server. Upon first send in any process, the connectivty-checker agent will send some UDP pings to the peer to ensure that we can reach it. If we can't, we'll abort the job with a nice show_help message. There's a lengthy comment in btl_usnic_connectivity.h explains the scheme and how it works. Reviewed by Dave Goodell. cmr=v1.7.5:ticket=trac:4253 This commit was SVN r30860. The following Trac tickets were found above: Ticket 4253 --> https://svn.open-mpi.org/trac/ompi/ticket/4253
2014-02-27 02:21:25 +04:00
CHECK(reg_int("connectivity_ack_timeout",
"Timeout, in milliseconds, while waiting for an ACK while verification connectivity between usNIC devices. If 0, the connectivity check is disabled (must be >=0).",
mca_btl_usnic_component.connectivity_ack_timeout,
&mca_btl_usnic_component.connectivity_ack_timeout,
REGINT_GE_ZERO, OPAL_INFO_LVL_3));
btl_usnic_mca.c: Increase default connectivity checker frequency In abusive MPI communication patterns, sending a UDP ping only once a second may not be sufficient -- all the UDP pings may be dropped. So increase the frequency of the pings to every quarter second, and allow more total pings to be sent. Total timeout time is still the same (10 seconds) -- we'll just now try 40 times (i.e., once every quarter second) as opposed to 10 times (i.e., once a second). Testing has shown that this frequency allows the connectivity checker to always succeed even in the many-to-one abusive communication patterns. cmr=v1.8.2:reviewer=dgoodell This commit was SVN r31602.
2014-05-02 15:06:18 +04:00
mca_btl_usnic_component.connectivity_num_retries = 40;
Add usnic connectivity-checking agent service. Basically: since usnic is a connectionless transport, we do not get OS-provided services "for free" that connection-oriented transports get, namely: "hey, I wasn't able to make a connection to peer X", and "hey, your connection to peer X has died." This connectivity-checker runs in a separate progress thread in the usnic BTL in local rank 0 on each server. Upon first send in any process, the connectivty-checker agent will send some UDP pings to the peer to ensure that we can reach it. If we can't, we'll abort the job with a nice show_help message. There's a lengthy comment in btl_usnic_connectivity.h explains the scheme and how it works. Reviewed by Dave Goodell. cmr=v1.7.5:ticket=trac:4253 This commit was SVN r30860. The following Trac tickets were found above: Ticket 4253 --> https://svn.open-mpi.org/trac/ompi/ticket/4253
2014-02-27 02:21:25 +04:00
CHECK(reg_int("connectivity_error_num_retries",
"Number of times to retry usNIC connectivity verification before aborting the MPI job (must be >0).",
mca_btl_usnic_component.connectivity_num_retries,
&mca_btl_usnic_component.connectivity_num_retries,
REGINT_GE_ONE, OPAL_INFO_LVL_3));
First commit of the Cisco usNIC BTL. This BTL accesses the Cisco usNIC Linux device via the Linux verbs API via Unreliable Datagram queue pairs. A few noteworthy points: * This BTL does most of its own fragmentation; it tells the PML that it has a very high max_send_size (much higher than the network MTU). * Since UD fragments are, by definition, unreliable, the usnic BTL handles all of its own reliability via a sliding window approach using the opal_hotel construct and many tricks stolen from the corpus of knowledge surrounding efficient TCP. * There is a fun PML latency-metric based optimization for NUMA awareness of short messages. * Note that this is ''not'' a generic UD verbs BTL; it is specific to the Cisco usNIC device. This commit was SVN r28879.
2013-07-20 02:13:58 +04:00
/* Register some synonyms to the ompi common verbs component */
ompi_common_verbs_mca_register(&mca_btl_usnic_component.super.btl_version);
return ret;
}
Копировать постоянную ссылку