[ofa-general] Re: [PATCH] opensm: preserve base lid routes
Al Chu
chu11 at llnl.gov
Tue Jun 17 09:25:45 PDT 2008
Hey Hal,
> Other than the average bandwidth numbers at the top which show PL >
> non PL, the ones in the table seem to be the other way around. Do
> those numbers indicate bandwidth ?
Sorry, the numbers for the mvapich/openmpi numbers are in units of time.
I think microseconds. So lower is better. The "degradation" columns
are the percent worse than LMC=0.
The key thing to see is that originally, when LMC > 0 and multiple paths
are not used, the numbers go up/it takes more time. When the
preserving-base-lids patch is put in, the LMC > 0 numbers are about the
same as the LMC=0 numbers.
Al
On Tue, 2008-06-17 at 07:51 -0700, Hal Rosenstock wrote:
> Al,
>
> On Mon, 2008-06-16 at 10:56 -0700, Al Chu wrote:
> > Hey Sasha,
> >
> > patch series seems to perform well. As a reminder mvapich 0.9.9 does
> > not how to deal with multiple lids, openMPI 1.2.6 does. The "preserve
> > lids" condition is labeled "PL".
> >
> > With LMC > 0, the preserve lids seems to do its job under mvapich 0.9.9,
> > and atleast maintains (sometimes increases) performance of openmpi 1.2.6
> > with LMC > 0. Numbers are attached.
>
> Maybe I'm not reading this right but here's what I see:
>
> Other than the average bandwidth numbers at the top which show PL > non
> PL, the ones in the table seem to be the other way around. Do those
> numbers indicate bandwidth ?
>
> Also, you mentioned that mvapich doesn't know how to deal with LMC yet
> the numbers seem to go up as LMC does.
>
> -- Hal
>
> > I'd give my thumbs up to the patch. Lets commit it.
> >
> > Al
> >
> > On Tue, 2008-06-10 at 05:59 +0300, Sasha Khapyorsky wrote:
> > > Basically this addresses the problem described by Al Chu in:
> > >
> > > http://lists.openfabrics.org/pipermail/general/2008-April/049132.html
> > >
> > > When base lid paths become completely disbalanced on a fabrics with
> > > lmc > 0.
> > >
> > > One feedback was from Yiftah Shahar:
> > >
> > > "I think that our requirements should be that even when you are working
> > > with LMC>0 then the base LID routing should not be affected.
> > > One way to achieve this goal is to first run the base-LID routing (so
> > > all base LID improvement will be also in LMC>0) and then start with the
> > > other LIDs as round-robbing starting from the base-lid-port + 1
> > > according current routing algorithm rules (keeping min-hop, up/down...)."
> > >
> > > We had some discussion with Al and Yiftah about this and considered that
> > > in addition to "pure" base lid paths preservation (which is good thing by
> > > itself) proposed method solves original lid disbalancing problem as well.
> > >
> > > This patch is implementation of the idea above.
> > >
> > > Signed-off-by: Sasha Khapyorsky <sashak at voltaire.com>
> > > ---
> > > opensm/include/opensm/osm_switch.h | 4 +
> > > opensm/opensm/osm_dump.c | 3 +-
> > > opensm/opensm/osm_switch.c | 8 ++-
> > > opensm/opensm/osm_ucast_mgr.c | 163 ++++++++++++++++--------------------
> > > 4 files changed, 86 insertions(+), 92 deletions(-)
> > >
> > > diff --git a/opensm/include/opensm/osm_switch.h b/opensm/include/opensm/osm_switch.h
> > > index 0e9c5fa..c1521a6 100644
> > > --- a/opensm/include/opensm/osm_switch.h
> > > +++ b/opensm/include/opensm/osm_switch.h
> > > @@ -981,6 +981,7 @@ uint8_t
> > > osm_switch_recommend_path(IN const osm_switch_t * const p_sw,
> > > IN osm_port_t * p_port,
> > > IN const uint16_t lid_ho,
> > > + IN unsigned start_from,
> > > IN const boolean_t ignore_existing,
> > > IN const boolean_t dor);
> > > /*
> > > @@ -995,6 +996,9 @@ osm_switch_recommend_path(IN const osm_switch_t * const p_sw,
> > > * lid_ho
> > > * [in] LID value (host order) for which to get a path advisory.
> > > *
> > > +* start_from
> > > +* [in] Port number from where to start balance counting.
> > > +*
> > > * ignore_existing
> > > * [in] Set to cause the switch to choose the optimal route
> > > * regardless of existing paths.
> > > diff --git a/opensm/opensm/osm_dump.c b/opensm/opensm/osm_dump.c
> > > index b96984b..60c6d25 100644
> > > --- a/opensm/opensm/osm_dump.c
> > > +++ b/opensm/opensm/osm_dump.c
> > > @@ -218,7 +218,8 @@ static void dump_ucast_routes(cl_map_item_t *p_map_item, FILE *file, void *cxt)
> > > else {
> > > /* No LMC Optimization */
> > > best_port = osm_switch_recommend_path(p_sw, p_port,
> > > - lid_ho, TRUE, dor);
> > > + lid_ho, 1, TRUE,
> > > + dor);
> > > fprintf(file, "No %u hop path possible via port %u!",
> > > best_hops, best_port);
> > > }
> > > diff --git a/opensm/opensm/osm_switch.c b/opensm/opensm/osm_switch.c
> > > index 58936e3..a9d13c8 100644
> > > --- a/opensm/opensm/osm_switch.c
> > > +++ b/opensm/opensm/osm_switch.c
> > > @@ -274,6 +274,7 @@ uint8_t
> > > osm_switch_recommend_path(IN const osm_switch_t * const p_sw,
> > > IN osm_port_t * p_port,
> > > IN const uint16_t lid_ho,
> > > + IN unsigned start_from,
> > > IN const boolean_t ignore_existing,
> > > IN const boolean_t dor)
> > > {
> > > @@ -294,6 +295,7 @@ osm_switch_recommend_path(IN const osm_switch_t * const p_sw,
> > > uint8_t port_num;
> > > uint8_t num_ports;
> > > uint32_t least_paths = 0xFFFFFFFF;
> > > + unsigned i;
> > > /*
> > > The follwing will track the least paths if the
> > > route should go through a new system/node
> > > @@ -397,8 +399,10 @@ osm_switch_recommend_path(IN const osm_switch_t * const p_sw,
> > > */
> > >
> > > /* port number starts with one and num_ports is 1 + num phys ports */
> > > - for (port_num = 1; port_num < num_ports; port_num++) {
> > > - if (osm_switch_get_hop_count(p_sw, base_lid, port_num) !=
> > > + for (i = start_from; i < start_from + num_ports; i++) {
> > > + port_num = i%num_ports;
> > > + if (!port_num ||
> > > + osm_switch_get_hop_count(p_sw, base_lid, port_num) !=
> > > least_hops)
> > > continue;
> > >
> > > diff --git a/opensm/opensm/osm_ucast_mgr.c b/opensm/opensm/osm_ucast_mgr.c
> > > index c073037..2aae6d5 100644
> > > --- a/opensm/opensm/osm_ucast_mgr.c
> > > +++ b/opensm/opensm/osm_ucast_mgr.c
> > > @@ -208,7 +208,8 @@ find_and_add_remote_sys(osm_switch_t *sw, uint8_t port,
> > > static void
> > > __osm_ucast_mgr_process_port(IN osm_ucast_mgr_t * const p_mgr,
> > > IN osm_switch_t * const p_sw,
> > > - IN osm_port_t * const p_port)
> > > + IN osm_port_t * const p_port,
> > > + IN unsigned lid_offset)
> > > {
> > > uint16_t min_lid_ho;
> > > uint16_t max_lid_ho;
> > > @@ -217,19 +218,14 @@ __osm_ucast_mgr_process_port(IN osm_ucast_mgr_t * const p_mgr,
> > > boolean_t is_ignored_by_port_prof;
> > > ib_net64_t node_guid;
> > > struct osm_routing_engine *p_routing_eng;
> > > - /*
> > > - The following are temporary structures that will aid
> > > - in providing better routing in LMC > 0 situations
> > > - */
> > > - uint16_t lids_per_port = 1 << p_mgr->p_subn->opt.lmc;
> > > - struct osm_remote_node *p_remote_guid_used = NULL;
> > > + unsigned start_from = 1;
> > >
> > > OSM_LOG_ENTER(p_mgr->p_log);
> > >
> > > osm_port_get_lid_range_ho(p_port, &min_lid_ho, &max_lid_ho);
> > >
> > > - /* If the lids are zero - then there was some problem with the initialization.
> > > - Don't handle this port. */
> > > + /* If the lids are zero - then there was some problem with
> > > + * the initialization. Don't handle this port. */
> > > if (min_lid_ho == 0 || max_lid_ho == 0) {
> > > OSM_LOG(p_mgr->p_log, OSM_LOG_ERROR, "ERR 3A04: "
> > > "Port 0x%" PRIx64 " has LID 0. An initialization "
> > > @@ -238,16 +234,22 @@ __osm_ucast_mgr_process_port(IN osm_ucast_mgr_t * const p_mgr,
> > > goto Exit;
> > > }
> > >
> > > - if (osm_log_is_active(p_mgr->p_log, OSM_LOG_DEBUG)) {
> > > + lid_ho = min_lid_ho + lid_offset;
> > > +
> > > + if (lid_ho > max_lid_ho)
> > > + goto Exit;
> > > +
> > > + if (lid_offset)
> > > + /* ignore potential overflow - it is handled in osm_switch.c */
> > > + start_from = osm_switch_get_port_by_lid(p_sw, lid_ho - 1) + 1;
> > > +
> > > + if (osm_log_is_active(p_mgr->p_log, OSM_LOG_DEBUG))
> > > OSM_LOG(p_mgr->p_log, OSM_LOG_DEBUG,
> > > - "Processing port 0x%" PRIx64 ", LIDs [0x%X,0x%X]\n",
> > > - cl_ntoh64(osm_port_get_guid(p_port)),
> > > + "Processing port 0x%" PRIx64 ", LID %u [0x%X,0x%X]\n",
> > > + cl_ntoh64(osm_port_get_guid(p_port)), lid_ho,
> > > min_lid_ho, max_lid_ho);
> > > - }
> > >
> > > - /*
> > > - TO DO - This should be runtime error, not a CL_ASSERT()
> > > - */
> > > + /* TODO - This should be runtime error, not a CL_ASSERT() */
> > > CL_ASSERT(max_lid_ho < osm_switch_get_fwd_tbl_size(p_sw));
> > >
> > > node_guid = osm_node_get_node_guid(p_sw->p_node);
> > > @@ -260,80 +262,62 @@ __osm_ucast_mgr_process_port(IN osm_ucast_mgr_t * const p_mgr,
> > > how best to distribute the LID range across the ports
> > > that can reach those LIDs.
> > > */
> > > - for (lid_ho = min_lid_ho; lid_ho <= max_lid_ho; lid_ho++) {
> > > - /* Use the enhanced algorithm only for LMC > 0 */
> > > - if (lids_per_port > 1) {
> > > - port = osm_switch_recommend_path(p_sw, p_port, lid_ho,
> > > - p_mgr->p_subn->
> > > - ignore_existing_lfts,
> > > - p_mgr->is_dor);
> > > - if (port > 0 && port != OSM_NO_PATH && p_port->priv)
> > > - p_remote_guid_used =
> > > - find_and_add_remote_sys(p_sw, port,
> > > - p_port->priv);
> > > + port = osm_switch_recommend_path(p_sw, p_port, lid_ho, start_from,
> > > + p_mgr->p_subn->ignore_existing_lfts,
> > > + p_mgr->is_dor);
> > > +
> > > + if (port == OSM_NO_PATH) {
> > > + /* do not try to overwrite the ppro of non existing port ... */
> > > + is_ignored_by_port_prof = TRUE;
> > > +
> > > + /* Up/Down routing can cause unreachable routes between some
> > > + switches so we do not report that as an error in that case */
> > > + if (!p_routing_eng->build_lid_matrices) {
> > > + OSM_LOG(p_mgr->p_log, OSM_LOG_ERROR, "ERR 3A08: "
> > > + "No path to get to LID 0x%X from switch 0x%"
> > > + PRIx64 "\n", lid_ho, cl_ntoh64(node_guid));
> > > + /* trigger a new sweep - try again ... */
> > > + p_mgr->p_subn->subnet_initialization_error = TRUE;
> > > } else
> > > - port = osm_switch_recommend_path(p_sw, p_port, lid_ho,
> > > - p_mgr->p_subn->
> > > - ignore_existing_lfts,
> > > - p_mgr->is_dor);
> > > + OSM_LOG(p_mgr->p_log, OSM_LOG_DEBUG,
> > > + "No path to get to LID 0x%X from switch 0x%"
> > > + PRIx64 "\n", lid_ho, cl_ntoh64(node_guid));
> > > + } else {
> > > + OSM_LOG(p_mgr->p_log, OSM_LOG_DEBUG,
> > > + "Routing LID 0x%X to port 0x%X"
> > > + " for switch 0x%" PRIx64 "\n",
> > > + lid_ho, port, cl_ntoh64(node_guid));
> > >
> > > /*
> > > - There might be no path to the target
> > > + we would like to optionally ignore this port in equalization
> > > + as in the case of the Mellanox Anafa Internal PCI TCA port
> > > */
> > > - if (port == OSM_NO_PATH) {
> > > - /* do not try to overwrite the ppro of non existing port ... */
> > > - is_ignored_by_port_prof = TRUE;
> > > -
> > > - /* Up/Down routing can cause unreachable routes between some
> > > - switches so we do not report that as an error in that case */
> > > - if (!p_routing_eng->build_lid_matrices) {
> > > - OSM_LOG(p_mgr->p_log, OSM_LOG_ERROR, "ERR 3A08: "
> > > - "No path to get to LID 0x%X from switch 0x%"
> > > - PRIx64 "\n", lid_ho,
> > > - cl_ntoh64(node_guid));
> > > - /* trigger a new sweep - try again ... */
> > > - p_mgr->p_subn->subnet_initialization_error =
> > > - TRUE;
> > > - } else
> > > - OSM_LOG(p_mgr->p_log, OSM_LOG_DEBUG,
> > > - "No path to get to LID 0x%X from switch 0x%"
> > > - PRIx64 "\n", lid_ho,
> > > - cl_ntoh64(node_guid));
> > > - } else {
> > > - OSM_LOG(p_mgr->p_log, OSM_LOG_DEBUG,
> > > - "Routing LID 0x%X to port 0x%X"
> > > - " for switch 0x%" PRIx64 "\n",
> > > - lid_ho, port, cl_ntoh64(node_guid));
> > > -
> > > - /*
> > > - we would like to optionally ignore this port in equalization
> > > - as in the case of the Mellanox Anafa Internal PCI TCA port
> > > - */
> > > - is_ignored_by_port_prof =
> > > - osm_port_prof_is_ignored_port(p_mgr->p_subn,
> > > - node_guid, port);
> > > -
> > > - /*
> > > - We also would ignore this route if the target lid is of a switch
> > > - and the port_profile_switch_node is not TRUE
> > > - */
> > > - if (!p_mgr->p_subn->opt.port_profile_switch_nodes) {
> > > - is_ignored_by_port_prof |=
> > > - (osm_node_get_type(p_port->p_node) ==
> > > - IB_NODE_TYPE_SWITCH);
> > > - }
> > > - }
> > > + is_ignored_by_port_prof =
> > > + osm_port_prof_is_ignored_port(p_mgr->p_subn,
> > > + node_guid, port);
> > >
> > > /*
> > > - We have selected the port for this LID.
> > > - Write it to the forwarding tables.
> > > + We also would ignore this route if the target lid is of
> > > + a switch and the port_profile_switch_node is not TRUE
> > > */
> > > - p_mgr->lft_buf[lid_ho] = port;
> > > - if (!is_ignored_by_port_prof) {
> > > - osm_switch_count_path(p_sw, port);
> > > - if (p_remote_guid_used)
> > > - p_remote_guid_used->forwarded_to++;
> > > - }
> > > + if (!p_mgr->p_subn->opt.port_profile_switch_nodes)
> > > + is_ignored_by_port_prof |=
> > > + (osm_node_get_type(p_port->p_node) ==
> > > + IB_NODE_TYPE_SWITCH);
> > > + }
> > > +
> > > + /*
> > > + We have selected the port for this LID.
> > > + Write it to the forwarding tables.
> > > + */
> > > + p_mgr->lft_buf[lid_ho] = port;
> > > + if (!is_ignored_by_port_prof) {
> > > + struct osm_remote_node *rem_node_used;
> > > + osm_switch_count_path(p_sw, port);
> > > + if (port > 0 && p_port->priv &&
> > > + (rem_node_used = find_and_add_remote_sys(p_sw, port,
> > > + p_port->priv)))
> > > + rem_node_used->forwarded_to++;
> > > }
> > >
> > > Exit:
> > > @@ -512,6 +496,7 @@ __osm_ucast_mgr_process_tbl(IN cl_map_item_t * const p_map_item,
> > > osm_node_t *p_node;
> > > osm_port_t *p_port;
> > > const cl_qmap_t *p_port_tbl;
> > > + unsigned i, lids_per_port;
> > >
> > > OSM_LOG_ENTER(p_mgr->p_log);
> > >
> > > @@ -538,12 +523,12 @@ __osm_ucast_mgr_process_tbl(IN cl_map_item_t * const p_map_item,
> > > Iterate through every port setting LID routes for each
> > > port based on base LID and LMC value.
> > > */
> > > -
> > > - for (p_port = (osm_port_t *) cl_qmap_head(p_port_tbl);
> > > - p_port != (osm_port_t *) cl_qmap_end(p_port_tbl);
> > > - p_port = (osm_port_t *) cl_qmap_next(&p_port->map_item)) {
> > > - __osm_ucast_mgr_process_port(p_mgr, p_sw, p_port);
> > > - }
> > > + lids_per_port = 1 << p_mgr->p_subn->opt.lmc;
> > > + for (i = 0; i < lids_per_port; i++)
> > > + for (p_port = (osm_port_t *) cl_qmap_head(p_port_tbl);
> > > + p_port != (osm_port_t *) cl_qmap_end(p_port_tbl);
> > > + p_port = (osm_port_t *) cl_qmap_next(&p_port->map_item))
> > > + __osm_ucast_mgr_process_port(p_mgr, p_sw, p_port, i);
> > >
> > > osm_ucast_mgr_set_fwd_table(p_mgr, p_sw);
> > >
>
--
Albert Chu
chu11 at llnl.gov
925-422-5311
Computer Scientist
High Performance Systems Division
Lawrence Livermore National Laboratory
More information about the general
mailing list