OPENSM(8) | OpenIB Management | OPENSM(8) |
opensm - InfiniBand subnet manager and administration (SM/SA)
opensm [--version]] [-F | --config <file_name>] [-c(reate-config) <file_name>] [-g(uid) <GUID in hex>] [-l(mc) <LMC>] [-p(riority) <PRIORITY>] [--subnet_prefix <PREFIX in hex>] [--smkey <SM_Key>] [--sm_sl <SL number>] [-r(eassign_lids)] [-R <engine name(s)> | --routing_engine <engine name(s)>] [--do_mesh_analysis] [--lash_start_vl <vl number>] [--nue_max_num_vls <vl number>] [-A | --ucast_cache] [-z | --connect_roots] [-M <file name> | --lid_matrix_file <file name>] [-U <file name> | --lfts_file <file name>] [-S | --sadb_file <file name>] [-a | --root_guid_file <path to file>] [-u | --cn_guid_file <path to file>] [-G | --io_guid_file <path to file>] [--port-shifting] [--scatter-ports <random seed>] [-H | --max_reverse_hops <max reverse hops allowed>] [-X | --guid_routing_order_file <path to file>] [-m | --ids_guid_file <path to file>] [-o(nce)] [-s(weep) <interval>] [-t(imeout) <milliseconds>] [--retries <number>] [--maxsmps <number>] [--console [off | local | socket | loopback]] [--console-port <port>] [-i | --ignore_guids <equalize-ignore-guids-file>] [-w | --hop_weights_file <path to file>] [-O | --port_search_ordering_file <path to file>] [-O | --dimn_ports_file <path to file>] (DEPRECATED) [--dump_files_dir <directory-name>] [-f <log file path> | --log_file <log file path> ] [-L | --log_limit <size in MB>] [-e(rase_log_file)] [-P(config) <partition config file> ] [-N | --no_part_enforce] (DEPRECATED) [-Z | --part_enforce [both | in | out | off]] [-W | --allow_both_pkeys] [-Q | --qos [-Y | --qos_policy_file <file name>]] [--congestion-control] [--cckey <key>] [-y | --stay_on_fatal] [-B | --daemon] [-J | --pidfile <file_name>] [-I | --inactive] [--perfmgr] [--perfmgr_sweep_time_s <seconds>] [--prefix_routes_file <path>] [--consolidate_ipv6_snm_req] [--log_prefix <prefix text>] [--torus_config <path to file>] [-v(erbose)] [-V] [-D <flags>] [-d(ebug) <number>] [-h(elp)] [-?]
opensm is an InfiniBand compliant Subnet Manager and Administration, and runs on top of OpenIB.
opensm provides an implementation of an InfiniBand Subnet Manager and Administration. Such a software entity is required to run for in order to initialize the InfiniBand hardware (at least one per each InfiniBand subnet).
opensm also now contains an experimental version of a performance manager as well.
opensm defaults were designed to meet the common case usage on clusters with up to a few hundred nodes. Thus, in this default mode, opensm will scan the IB fabric, initialize it, and sweep occasionally for changes.
opensm attaches to a specific IB port on the local machine and configures only the fabric connected to it. (If the local machine has other IB ports, opensm will ignore the fabrics connected to those other ports). If no port is specified, it will select the first "best" available port.
opensm can present the available ports and prompt for a port number to attach to.
By default, the run is logged to two files: /var/log/messages and /var/log/opensm.log. The first file will register only general major events, whereas the second will include details of reported errors. All errors reported in this second file should be treated as indicators of IB fabric health issues. (Note that when a fatal and non-recoverable error occurs, opensm will exit.) Both log files should include the message "SUBNET UP" if opensm was able to setup the subnet correctly.
BIT LOG LEVEL ENABLED
---- -----------------
0x01 - ERROR (error messages)
0x02 - INFO (basic messages, low volume)
0x04 - VERBOSE (interesting stuff, moderate volume)
0x08 - DEBUG (diagnostic, high volume)
0x10 - FUNCS (function entry/exit, very high volume)
0x20 - FRAMES (dumps all SMP and GMP frames)
0x40 - ROUTING (dump FDB routing information)
0x80 - SYS (syslog at LOG_INFO level in addition to OpenSM logging)
Without -D, OpenSM defaults to ERROR + INFO (0x3). Specifying -D 0 disables all messages. Specifying -D 0xFF enables all messages (see -V). High verbosity levels may require increasing the transaction timeout with the -t option.
OPT Description
--- -----------------
-d0 - Ignore other SM nodes
-d1 - Force single threaded dispatching
-d2 - Force log flushing after each log message
-d3 - Disable multicast support
The following environment variables control opensm behavior:
OSM_TMP_DIR - controls the directory in which the temporary files generated by opensm are created. These files are: opensm-subnet.lst, opensm.fdbs, and opensm.mcfdbs. By default, this directory is /var/log. Note that --dump_files_dir command line option or dump_file_dir option in option/config file takes precedence over this environment variable.
OSM_CACHE_DIR - opensm stores certain data to the disk such that subsequent runs are consistent. The default directory used is /var/cache/opensm. The following files are included in it:
guid2lid - stores the LID range assigned to each GUID
guid2mkey - stores the MKey previously assigned to each GUID
neighbors - stores a map of the GUIDs at either end of each link
in the fabric
When opensm receives a HUP signal, it starts a new heavy sweep as if a trap was received or a topology change was found.
Also, SIGUSR1 can be used to trigger a reopen of /var/log/opensm.log for logrotate purposes.
The default name of OpenSM partitions configuration file is /etc/opensm/partitions.conf. The default may be changed by using the --Pconfig (-P) option with OpenSM.
The default partition will be created by OpenSM unconditionally even when partition configuration file does not exist or cannot be accessed.
The default partition has P_Key value 0x7fff. OpenSM´s port will always have full membership in default partition. All other end ports will have full membership if the partition configuration file is not found or cannot be accessed, or limited membership if the file exists and can be accessed but there is no rule for the Default partition.
Effectively, this amounts to the same as if one of the following rules below appear in the partition configuration file.
In the case of no rule for the Default partition:
Default=0x7fff : ALL=limited, SELF=full ;
In the case of no partition configuration file or file cannot be accessed:
Default=0x7fff : ALL=full ;
File Format
Comments:
Line content followed after ´#´ character is comment and ignored by parser.
General file format:
<Partition Definition>:[<newline>]<Partition Properties>;
Partition Definition:
[PartitionName][=PKey][,indx0][,ipoib_bc_flags][,defmember=full|limited]
PartitionName - string, will be used with logging. When
omitted, empty string will be used.
PKey - P_Key value for this partition. Only low 15
bits will be used. When omitted will be
autogenerated.
indx0 - indicates that this pkey should be inserted in
block 0 index 0.
ipoib_bc_flags - used to indicate/specify IPoIB capability of
this partition.
defmember=full|limited|both - specifies default membership for
port guid list. Default is limited.
ipoib_bc_flags:
ipoib_flag|[mgroup_flag]*
ipoib_flag:
ipoib - indicates that this partition may be used for
IPoIB, as a result the IPoIB broadcast group will
be created with the mgroup_flag flags given,
if any.
Partition Properties:
[<Port list>|<MCast Group>]* | <Port list>
Port list:
<Port Specifier>[,<Port Specifier>]
Port Specifier:
<PortGUID>[=[full|limited|both]]
PortGUID - GUID of partition member EndPort.
Hexadecimal numbers should start from
0x, decimal numbers are accepted too.
full, limited, - indicates full and/or limited membership for
both this port. When omitted (or unrecognized)
limited membership is assumed. Both
indicates both full and limited membership
for this port.
MCast Group:
mgid=gid[,mgroup_flag]*<newline>
- gid specified is verified to be a Multicast
address. IP groups are verified to match
the rate and mtu of the broadcast group.
The P_Key bits of the mgid for IP groups are
verified to either match the P_Key specified
in by "Partition Definition" or if they are
0x0000 the P_Key will be copied into those
bits.
mgroup_flag:
rate=<val> - specifies rate for this MC group
(default is 3 (10GBps))
mtu=<val> - specifies MTU for this MC group
(default is 4 (2048))
sl=<val> - specifies SL for this MC group
(default is 0)
scope=<val> - specifies scope for this MC group
(default is 2 (link local)). Multiple scope
settings are permitted for a partition.
NOTE: This overwrites the scope nibble of the
specified mgid. Furthermore specifying
multiple scope settings will result in
multiple MC groups being created.
Q_Key=<val> - specifies the Q_Key for this MC group
(default: 0x0b1b for IP groups, 0 for other
groups)
WARNING: changing this for the broadcast
group may break IPoIB on client
nodes!!
TClass=<val> - specifies tclass for this MC group
(default is 0)
FlowLabel=<val> - specifies FlowLabel for this MC group
(default is 0) NOTE: All mgroup_flag flags MUST be separated by comma
(,).
Note that values for rate, mtu, and scope, for both partitions and multicast groups, should be specified as defined in the IBTA specification (for example, mtu=4 for 2048).
There are several useful keywords for PortGUID definition:
- 'ALL' means all end ports in this subnet.
- 'ALL_CAS' means all Channel Adapter end ports in this subnet.
- 'ALL_SWITCHES' means all Switch end ports in this subnet.
- 'ALL_ROUTERS' means all Router end ports in this subnet.
- 'SELF' means subnet manager's port.
Empty list means no ports in this partition.
Notes:
White space is permitted between delimiters ('=', ',',':',';').
PartitionName does not need to be unique, PKey does need to be unique. If PKey is repeated then those partition configurations will be merged and first PartitionName will be used (see also next note).
It is possible to split partition configuration in more than one definition, but then PKey should be explicitly specified (otherwise different PKey values will be generated for those definitions).
Examples:
Default=0x7fff : ALL, SELF=full ;
Default=0x7fff : ALL, ALL_SWITCHES=full, SELF=full ;
NewPartition , ipoib : 0x123456=full, 0x3456789034=limi, 0x2134af2306 ;
YetAnotherOne = 0x300 : SELF=full ;
YetAnotherOne = 0x300 : ALL=limited ;
ShareIO = 0x80 , defmember=full : 0x123451, 0x123452;
# 0x123453, 0x123454 will be limited
ShareIO = 0x80 : 0x123453, 0x123454, 0x123455=full;
# 0x123456, 0x123457 will be limited
ShareIO = 0x80 : defmember=limited : 0x123456, 0x123457, 0x123458=full;
ShareIO = 0x80 , defmember=full : 0x123459, 0x12345a;
ShareIO = 0x80 , defmember=full : 0x12345b, 0x12345c=limited, 0x12345d;
# multicast groups added to default
Default=0x7fff,ipoib:
mgid=ff12:401b::0707,sl=1 # random IPv4 group
mgid=ff12:601b::16 # MLDv2-capable routers
mgid=ff12:401b::16 # IGMP
mgid=ff12:601b::2 # All routers
mgid=ff12::1,sl=1,Q_Key=0xDEADBEEF,rate=3,mtu=2 # random group
ALL=full;
Note:
The following rule is equivalent to how OpenSM used to run prior to the partition manager:
Default=0x7fff,ipoib:ALL=full;
There are a set of QoS related low-level configuration parameters. All these parameter names are prefixed by "qos_" string. Here is a full list of these parameters:
qos_max_vls - The maximum number of VLs that will be on the subnet
qos_high_limit - The limit of High Priority component of VL
Arbitration table (IBA 7.6.9)
qos_vlarb_low - Low priority VL Arbitration table (IBA 7.6.9)
template
qos_vlarb_high - High priority VL Arbitration table (IBA 7.6.9)
template
Both VL arbitration templates are pairs of
VL and weight
qos_sl2vl - SL2VL Mapping table (IBA 7.6.6) template. It is
a list of VLs corresponding to SLs 0-15 (Note
that VL15 used here means drop this SL)
Typical default values (hard-coded in OpenSM initialization) are:
qos_max_vls 15
qos_high_limit 0
qos_vlarb_low 0:0,1:4,2:4,3:4,4:4,5:4,6:4,7:4,8:4,9:4,10:4,11:4,12:4,13:4,14:4
qos_vlarb_high 0:4,1:0,2:0,3:0,4:0,5:0,6:0,7:0,8:0,9:0,10:0,11:0,12:0,13:0,14:0
The syntax is compatible with rest of OpenSM configuration options and values may be stored in OpenSM config file (cached options file).
In addition to the above, we may define separate QoS configuration parameters sets for various target types. As targets, we currently support CAs, routers, switch external ports, and switch's enhanced port 0. The names of such specialized parameters are prefixed by "qos_<type>_" string. Here is a full list of the currently supported sets:
qos_ca_ - QoS configuration parameters set for CAs.
qos_rtr_ - parameters set for routers.
qos_sw0_ - parameters set for switches' port 0.
qos_swe_ - parameters set for switches' external ports.
Examples:
qos_sw0_max_vls=2
qos_ca_sl2vl=0,1,2,3,5,5,5,12,12,0,
qos_swe_high_limit=0
Prefix routes control how the SA responds to path record queries for off-subnet DGIDs. By default, the SA fails such queries. Note that IBA does not specify how the SA should obtain off-subnet path record information. The prefix routes configuration is meant as a stop-gap until the specification is completed.
Each line in the configuration file is a 64-bit prefix followed by a 64-bit GUID, separated by white space. The GUID specifies the router port on the local subnet that will handle the prefix. Blank lines are ignored, as is anything between a # character and the end of the line. The prefix and GUID are both in hex, the leading 0x is optional. Either, or both, can be wild-carded by specifying an asterisk instead of an explicit prefix or GUID.
When responding to a path record query for an off-subnet DGID, opensm searches for the first prefix match in the configuration file. Therefore, the order of the lines in the configuration file is important: a wild-carded prefix at the beginning of the configuration file renders all subsequent lines useless. If there is no match, then opensm fails the query. It is legal to repeat prefixes in the configuration file, opensm will return the path to the first available matching router. A configuration file with a single line where both prefix and GUID are wild-carded means that a path record query specifying any off-subnet DGID should return a path to the first available router. This configuration yields the same behavior formerly achieved by compiling opensm with -DROUTER_EXP which has been obsoleted.
OpenSM supports configuring a single management key (MKey) for use across the subnet.
The following configuration options are available:
m_key - the 64-bit MKey to be used on the subnet
(IBA 14.2.4)
m_key_protection_level - the numeric value of the MKey ProtectBits
(IBA 14.2.4.1)
m_key_lease_period - the number of seconds a CA will wait for a
response from the SM before resetting the
protection level to 0 (IBA 14.2.4.2).
OpenSM will configure all ports with the MKey specified by m_key, defaulting to a value of 0. A m_key value of 0 disables MKey protection on the subnet. Switches and HCAs with a non-zero MKey will not accept requests to change their configuration unless the request includes the proper MKey.
MKey Protection Levels
MKey protection levels modify how switches and CAs respond to SMPs lacking a valid MKey. OpenSM will configure each port's ProtectBits to support the level defined by the m_key_protection_level parameter. If no parameter is specified, OpenSM defaults to operating at protection level 0.
There are currently 4 protection levels defined by the IBA:
0 - Queries return valid data, including MKey. Configuration changes
are not allowed unless the request contains a valid MKey.
1 - Like level 0, but the MKey is set to 0 (0x00000000) in queries,
unless the request contains a valid MKey.
2 - Neither queries nor configuration changes are allowed, unless the
request contains a valid MKey.
3 - Identical to 2. Maintained for backwards compatibility.
MKey Lease Period
InfiniBand supports a MKey lease timeout, which is intended to allow administrators or a new SM to recover/reset lost MKeys on a fabric.
If MKeys are enabled on the subnet and a switch or CA receives a request that requires a valid MKey but does not contain one, it warns the SM by sending a trap (Bad M_Key, Trap 256). If the MKey lease period is non-zero, it also starts a countdown timer for the time specified by the lease period. If a SM (or other agent) responds with the correct MKey, the timer is stopped and reset. Should the timer reach zero, the switch or CA will reset its MKey protection level to 0, exposing the MKey and allowing recovery.
OpenSM will initialize all ports to use a mkey lease period of the number of seconds specified in the config file. If no mkey_lease_period is specified, a default of 0 will be used.
OpenSM normally quickly responds to all Bad_M_Key traps, resetting the lease timers. Additionally, OpenSM's subnet sweeps will also cancel any running timers. For maximum protection against accidentally-exposed MKeys, the MKey lease time should be a few multiples of the subnet sweep time. If OpenSM detects at startup that your sweep interval is greater than your MKey lease period, it will reset the lease period to be greater than the sweep interval. Similarly, if sweeping is disabled at startup, it will be re-enabled with an interval less than the Mkey lease period.
If OpenSM is required to recover a subnet for which it is missing mkeys, it must do so one switch level at a time. As such, the total time to recover the subnet may be as long as the mkey lease period multiplied by the maximum number of hops between the SM and an endpoint, plus one.
MKey Effects on Diagnostic Utilities
Setting a MKey may have a detrimental effect on diagnostic software run on the subnet, unless your diagnostic software is able to retrieve MKeys from the SA or can be explicitly configured with the proper MKey. This is particularly true at protection level 2, where CAs will ignore queries for management information that do not contain the proper MKey.
OpenSM now offers ten routing engines:
1. Min Hop Algorithm - based on the minimum hops to each node where the path length is optimized.
2. UPDN Unicast routing algorithm - also based on the minimum hops to each node, but it is constrained to ranking rules. This algorithm should be chosen if the subnet is not a pure Fat Tree, and deadlock may occur due to a loop in the subnet.
3. DNUP Unicast routing algorithm - similar to UPDN but allows routing in fabrics which have some CA nodes attached closer to the roots than some switch nodes.
4. Fat Tree Unicast routing algorithm - this algorithm optimizes routing for congestion-free "shift" communication pattern. It should be chosen if a subnet is a symmetrical or almost symmetrical fat-tree of various types, not just K-ary-N-Trees: non-constant K, not fully staffed, any Constant Bisectional Bandwidth (CBB) ratio. Similar to UPDN, Fat Tree routing is constrained to ranking rules.
5. LASH unicast routing algorithm - uses InfiniBand virtual layers (SL) to provide deadlock-free shortest-path routing while also distributing the paths between layers. LASH is an alternative deadlock-free topology-agnostic routing algorithm to the non-minimal UPDN algorithm avoiding the use of a potentially congested root node.
6. DOR Unicast routing algorithm - based on the Min Hop algorithm, but avoids port equalization except for redundant links between the same two switches. This provides deadlock free routes for hypercubes when the fabric is cabled as a hypercube and for meshes when cabled as a mesh (see details below).
7. Torus-2QoS unicast routing algorithm - a DOR-based routing algorithm specialized for 2D/3D torus topologies. Torus-2QoS provides deadlock-free routing while supporting two quality of service (QoS) levels. In addition it is able to route around multiple failed fabric links or a single failed fabric switch without introducing deadlocks, and without changing path SL values granted before the failure.
8. DFSSSP unicast routing algorithm - a deadlock-free single-source-shortest-path routing, which uses the SSSP algorithm (see algorithm 9.) as the base to optimize link utilization and uses InfiniBand virtual lanes (SL) to provide deadlock-freedom.
9. SSSP unicast routing algorithm - a single-source-shortest-path routing algorithm, which globally balances the number of routes per link to optimize link utilization. This routing algorithm has no restrictions in terms of the underlying topology.
10. Nue unicast routing algorithm - a 100%-applicable and deadlock-free routing which can be used for any arbitrary or faulty network topology and any number of virtual lanes (this includes the absence of VLs as well). Paths are globally balanced w.r.t the number of routes per link, and are kept as short as possible while enforcing deadlock-freedom within the VL constraint.
OpenSM also supports a file method which can load routes from a table. See ´Modular Routing Engine´ for more information on this.
The basic routing algorithm is comprised of two stages:
1. MinHop matrix calculation
How many hops are required to get from each port to each LID ?
The algorithm to fill these tables is different if you run standard (min hop)
or Up/Down.
For standard routing, a "relaxation" algorithm is used to propagate
min hop from every destination LID through neighbor switches
For Up/Down routing, a BFS from every target is used. The BFS tracks link
direction (up or down) and avoid steps that will perform up after a down
step was used.
2. Once MinHop matrices exist, each switch is visited and for each
target LID a decision is made as to what port should be used to get to that
LID.
This step is common to standard and Up/Down routing. Each port has a counter
counting the number of target LIDs going through it.
When there are multiple alternative ports with same MinHop to a LID, the one
with less previously assigned LIDs is selected.
If LMC > 0, more checks are added: Within each group of LIDs assigned to
same target port,
a. use only ports which have same MinHop
b. first prefer the ones that go to different systemImageGuid (then the
previous LID of the same LMC group)
c. if none - prefer those which go through another NodeGuid
d. fall back to the number of paths method (if all go to same node).
Effect of Topology Changes
OpenSM will preserve existing routing in any case where there is no change in the fabric switches unless the -r (--reassign_lids) option is specified.
-r
--reassign_lids
This option causes OpenSM to reassign LIDs to all
end nodes. Specifying -r on a running subnet
may disrupt subnet traffic.
Without -r, OpenSM attempts to preserve existing
LID assignments resolving multiple use of same LID.
If a link is added or removed, OpenSM does not recalculate the routes that do not have to change. A route has to change if the port is no longer UP or no longer the MinHop. When routing changes are performed, the same algorithm for balancing the routes is invoked.
In the case of using the file based routing, any topology changes are currently ignored The 'file' routing engine just loads the LFTs from the file specified, with no reaction to real topology. Obviously, this will not be able to recheck LIDs (by GUID) for disconnected nodes, and LFTs for non-existent switches will be skipped. Multicast is not affected by 'file' routing engine (this uses min hop tables).
Min Hop Algorithm
The Min Hop algorithm is invoked by default if no routing algorithm is specified. It can also be invoked by specifying '-R minhop'.
The Min Hop algorithm is divided into two stages: computation of min-hop tables on every switch and LFT output port assignment. Link subscription is also equalized with the ability to override based on port GUID. The latter is supplied by:
-i <equalize-ignore-guids-file>
--ignore_guids <equalize-ignore-guids-file>
This option provides the means to define a set of ports
(by guid) that will be ignored by the link load
equalization algorithm. Note that only endports (CA,
switch port 0, and router ports) and not switch external
ports are supported.
LMC awareness routes based on (remote) system or switch basis.
Purpose of UPDN Algorithm
The UPDN algorithm is designed to prevent deadlocks from occurring in loops of the subnet. A loop-deadlock is a situation in which it is no longer possible to send data between any two hosts connected through the loop. As such, the UPDN routing algorithm should be used if the subnet is not a pure Fat Tree, and one of its loops may experience a deadlock (due, for example, to high pressure).
The UPDN algorithm is based on the following main stages:
1. Auto-detect root nodes - based on the CA hop length from any switch in the subnet, a statistical histogram is built for each switch (hop num vs number of occurrences). If the histogram reflects a specific column (higher than others) for a certain node, then it is marked as a root node. Since the algorithm is statistical, it may not find any root nodes. The list of the root nodes found by this auto-detect stage is used by the ranking process stage.
Note 1: The user can override the node list manually.
Note 2: If this stage cannot find any root nodes, and the user did
not specify a guid list file, OpenSM defaults back to the
Min Hop routing algorithm.
2. Ranking process - All root switch nodes (found in stage 1) are assigned a rank of 0. Using the BFS algorithm, the rest of the switch nodes in the subnet are ranked incrementally. This ranking aids in the process of enforcing rules that ensure loop-free paths.
3. Min Hop Table setting - after ranking is done, a BFS algorithm is run from each (CA or switch) node in the subnet. During the BFS process, the FDB table of each switch node traversed by BFS is updated, in reference to the starting node, based on the ranking rules and guid values.
At the end of the process, the updated FDB tables ensure loop-free paths through the subnet.
Note: Up/Down routing does not allow LID routing communication between switches that are located inside spine "switch systems". The reason is that there is no way to allow a LID route between them that does not break the Up/Down rule. One ramification of this is that you cannot run SM on switches other than the leaf switches of the fabric.
UPDN Algorithm Usage
Activation through OpenSM
Use '-R updn' option (instead of old '-u') to activate the UPDN algorithm. Use '-a <root_guid_file>' for adding an UPDN guid file that contains the root nodes for ranking. If the `-a' option is not used, OpenSM uses its auto-detect root nodes algorithm.
Notes on the guid list file:
1. A valid guid file specifies one guid in each line. Lines with
an invalid format will be discarded.
2. The user should specify the root switch guids. However, it is also possible
to specify CA guids; OpenSM will use the guid of the switch (if it exists)
that connects the CA to the subnet as a root node.
Purpose of DNUP Algorithm
The DNUP algorithm is designed to serve a similar purpose to UPDN. However it is intended to work in network topologies which are unsuited to UPDN due to nodes being connected closer to the roots than some of the switches. An example would be a fabric which contains nodes and uplinks connected to the same switch. The operation of DNUP is the same as UPDN with the exception of the ranking process. In DNUP all switch nodes are ranked based solely on their distance from CA Nodes, all switch nodes directly connected to at least one CA are assigned a value of 1 all other switch nodes are assigned a value of one more than the minimum rank of all neighbor switch nodes.
Fat-tree Routing Algorithm
The fat-tree algorithm optimizes routing for "shift" communication pattern. It should be chosen if a subnet is a symmetrical or almost symmetrical fat-tree of various types. It supports not just K-ary-N-Trees, by handling for non-constant K, cases where not all leafs (CAs) are present, any CBB ratio. As in UPDN, fat-tree also prevents credit-loop-deadlocks.
If the root guid file is not provided ('-a' or '--root_guid_file'
options), the topology has to be pure fat-tree that complies with the
following rules:
- Tree rank should be between two and eight (inclusively)
- Switches of the same rank should have the same number
of UP-going port groups*, unless they are root switches,
in which case the shouldn't have UP-going ports at all.
- Switches of the same rank should have the same number
of DOWN-going port groups, unless they are leaf switches.
- Switches of the same rank should have the same number
of ports in each UP-going port group.
- Switches of the same rank should have the same number
of ports in each DOWN-going port group.
- All the CAs have to be at the same tree level (rank).
If the root guid file is provided, the topology doesn't have to be
pure fat-tree, and it should only comply with the following rules:
- Tree rank should be between two and eight (inclusively)
- All the Compute Nodes** have to be at the same tree level (rank).
Note that non-compute node CAs are allowed here to be at different
tree ranks.
* ports that are connected to the same remote switch are referenced as ´port group´.
** list of compute nodes (CNs) can be specified by ´-u´ or ´--cn_guid_file´ OpenSM options.
Topologies that do not comply cause a fallback to min hop routing. Note that this can also occur on link failures which cause the topology to no longer be "pure" fat-tree.
Note that although fat-tree algorithm supports trees with non-integer CBB ratio, the routing will not be as balanced as in case of integer CBB ratio. In addition to this, although the algorithm allows leaf switches to have any number of CAs, the closer the tree is to be fully populated, the more effective the "shift" communication pattern will be. In general, even if the root list is provided, the closer the topology to a pure and symmetrical fat-tree, the more optimal the routing will be.
The algorithm also dumps compute node ordering file (opensm-ftree-ca-order.dump) in the same directory where the OpenSM log resides. This ordering file provides the CN order that may be used to create efficient communication pattern, that will match the routing tables.
Routing between non-CN nodes
The use of the cn_guid_file option allows non-CN nodes to be located on different levels in the fat tree. In such case, it is not guaranteed that the Fat Tree algorithm will route between two non-CN nodes. To solve this problem, a list of non-CN nodes can be specified by ´-G´ or ´--io_guid_file´ option. Theses nodes will be allowed to use switches the wrong way round a specific number of times (specified by ´-H´ or ´--max_reverse_hops´. With the proper max_reverse_hops and io_guid_file values, you can ensure full connectivity in the Fat Tree.
Please note that using max_reverse_hops creates routes that use the switch in a counter-stream way. This option should never be used to connect nodes with high bandwidth traffic between them ! It should only be used to allow connectivity for HA purposes or similar. Also having routes the other way around can in theory cause credit loops.
Use these options with extreme care !
Activation through OpenSM
Use '-R ftree' option to activate the fat-tree algorithm. Use '-a <root_guid_file>' to provide root nodes for ranking. If the `-a' option is not used, routing algorithm will detect roots automatically. Use '-u <root_cn_file>' to provide the list of compute nodes. If the `-u' option is not used, all the CAs are considered as compute nodes.
Note: LMC > 0 is not supported by fat-tree routing. If this is specified, the default routing algorithm is invoked instead.
LASH Routing Algorithm
LASH is an acronym for LAyered SHortest Path Routing. It is a deterministic shortest path routing algorithm that enables topology agnostic deadlock-free routing within communication networks.
When computing the routing function, LASH analyzes the network topology for the shortest-path routes between all pairs of sources / destinations and groups these paths into virtual layers in such a way as to avoid deadlock.
Note LASH analyzes routes and ensures deadlock freedom between switch pairs. The link from HCA between and switch does not need virtual layers as deadlock will not arise between switch and HCA.
In more detail, the algorithm works as follows:
1) LASH determines the shortest-path between all pairs of source / destination switches. Note, LASH ensures the same SL is used for all SRC/DST - DST/SRC pairs and there is no guarantee that the return path for a given DST/SRC will be the reverse of the route SRC/DST.
2) LASH then begins an SL assignment process where a route is assigned to a layer (SL) if the addition of that route does not cause deadlock within that layer. This is achieved by maintaining and analysing a channel dependency graph for each layer. Once the potential addition of a path could lead to deadlock, LASH opens a new layer and continues the process.
3) Once this stage has been completed, it is highly likely that the first layers processed will contain more paths than the latter ones. To better balance the use of layers, LASH moves paths from one layer to another so that the number of paths in each layer averages out.
Note, the implementation of LASH in opensm attempts to use as few layers as possible. This number can be less than the number of actual layers available.
In general LASH is a very flexible algorithm. It can, for example, reduce to Dimension Order Routing in certain topologies, it is topology agnostic and fares well in the face of faults.
It has been shown that for both regular and irregular topologies, LASH outperforms Up/Down. The reason for this is that LASH distributes the traffic more evenly through a network, avoiding the bottleneck issues related to a root node and always routes shortest-path.
The algorithm was developed by Simula Research Laboratory.
Use '-R lash -Q ' option to activate the LASH algorithm.
Note: QoS support has to be turned on in order that SL/VL mappings are used.
Note: LMC > 0 is not supported by the LASH routing. If this is specified, the default routing algorithm is invoked instead.
For open regular cartesian meshes the DOR algorithm is the ideal routing algorithm. For toroidal meshes on the other hand there are routing loops that can cause deadlocks. LASH can be used to route these cases. The performance of LASH can be improved by preconditioning the mesh in cases where there are multiple links connecting switches and also in cases where the switches are not cabled consistently. An option exists for LASH to do this. To invoke this use '-R lash -Q --do_mesh_analysis'. This will add an additional phase that analyses the mesh to try to determine the dimension and size of a mesh. If it determines that the mesh looks like an open or closed cartesian mesh it reorders the ports in dimension order before the rest of the LASH algorithm runs.
DOR Routing Algorithm
The Dimension Order Routing algorithm is based on the Min Hop algorithm and so uses shortest paths. Instead of spreading traffic out across different paths with the same shortest distance, it chooses among the available shortest paths based on an ordering of dimensions. Each port must be consistently cabled to represent a hypercube dimension or a mesh dimension. Alternatively, the -O option can be used to assign a custom mapping between the ports on a given switch, and the associated dimension. Paths are grown from a destination back to a source using the lowest dimension (port) of available paths at each step. This provides the ordering necessary to avoid deadlock. When there are multiple links between any two switches, they still represent only one dimension and traffic is balanced across them unless port equalization is turned off. In the case of hypercubes, the same port must be used throughout the fabric to represent the hypercube dimension and match on both ends of the cable, or the -O option used to accomplish the alignment. In the case of meshes, the dimension should consistently use the same pair of ports, one port on one end of the cable, and the other port on the other end, continuing along the mesh dimension, or the -O option used as an override.
Use '-R dor' option to activate the DOR algorithm.
DFSSSP and SSSP Routing Algorithm
The (Deadlock-Free) Single-Source-Shortest-Path routing algorithm is designed to optimize link utilization thru global balancing of routes, while supporting arbitrary topologies. The DFSSSP routing algorithm uses InfiniBand virtual lanes (SL) to provide deadlock-freedom.
The DFSSSP algorithm consists of five major steps:
1) It discovers the subnet and models the subnet as a directed multigraph in
which each node represents a node of the physical network and each edge
represents one direction of the full-duplex links used to connect the nodes.
2) A loop, which iterates over all CA and switches of the subnet, will perform
three steps to generate the linear forwarding tables for each switch:
2.1) use Dijkstra's algorithm to find the shortest path from all nodes to the
current selected destination;
2.2) update the edge weights in the graph, i.e. add the number of routes,
which use a link to reach the destination, to the link/edge;
2.3) update the LFT of each switch with the outgoing port which was used in
the current step to route the traffic to the destination node.
3) After the number of available virtual lanes or layers in the subnet is
detected and a channel dependency graph is initialized for each layer, the
algorithm will put each possible route of the subnet into the first layer.
4) A loop iterates over all channel dependency graphs (CDG) and performs the
following substeps:
4.1) search for a cycle in the current CDG;
4.2) when a cycle is found, i.e. a possible deadlock is present, one edge is
selected and all routes, which induced this edge, are moved to the
"next higher" virtual layer (CDG[i+1]);
4.3) the cycle search is continued until all cycles are broken and routes are
moved "up".
5) When the number of needed layers does not exceeds the number of available
SL/VL to remove all cycles in all CDGs, the routing is deadlock-free and an
relation table is generated, which contains the assignment of routes from
source to destination to a SL
Note on SSSP:
This algorithm does not perform the steps 3)-5) and can not be considered to
be deadlock-free for all topologies. But on the one hand, you can choose
this algorithm for really large networks (5,000+ CAs and deadlock-free by
design) to reduce the runtime of the algorithm. On the other hand, you might
use the SSSP routing algorithm as an alternative, when all deadlock-free
routing algorithms fail to route the network for whatever reason. In the
last case, SSSP was designed to deliver an equal or higher bandwidth due to
better congestion avoidance than the Min Hop routing algorithm.
Notes for usage:
a) running DFSSSP: '-R dfsssp -Q'
a.1) QoS has to be configured to equally spread the load on the available SL
or virtual lanes
a.2) applications must perform a path record query to get path SL for each
route, which the application will use to transmit packages
b) running SSSP: '-R sssp'
c) both algorithms support LMC > 0
Hints for optimizing I/O traffic:
Having more nodes (I/O and compute) connected to a switch than incoming links
can result in a 'bad' routing of the I/O traffic as long as (DF)SSSP routing
is not aware of the dedicated I/O nodes, i.e., in the following network
configuration CN1-CN3 might send all I/O traffic via Link2 to IO1,IO2:
CN1 Link1 IO1
\ /----\ /
CN2 -- Switch1 Switch2 -- CN4
/ \----/ \
CN3 Link2 IO2
To prevent this from happening (DF)SSSP can use both the compute
node guid file and the I/O guid file specified by the ´-u´ or
´--cn_guid_file´ and ´-G´ or
´--io_guid_file´ options (similar to the Fat-Tree routing).
This ensures that traffic towards compute nodes and I/O nodes is balanced
separately and therefore distributed as much as possible across the
available links. Port GUIDs, as listed by ibstat, must be specified (not
Node GUIDs).
The priority for the optimization is as follows:
compute nodes -> I/O nodes -> other nodes
Possible use case scenarios:
a) neither ´-u´ nor ´-G´ are specified: all nodes
a treated as ´other nodes´ and therefore balanced equally;
b) ´-G´ is specified: traffic towards I/O nodes will be balanced
optimally;
c) the system has three node types, such as login/admin, compute and I/O, but
the balancing focus should be I/O, then one has to use ´-u´
and ´-G´ with I/O guids listed in cn_guid_file and compute
node guids listed in io_guid_file;
d) ...
Torus-2QoS Routing Algorithm
Torus-2QoS is routing algorithm designed for large-scale 2D/3D torus fabrics; see torus-2QoS(8) for full documentation.
Use '-R torus-2QoS -Q' or '-R torus-2QoS,no_fallback -Q' to activate the torus-2QoS algorithm.
Nue Routing Algorithm
Use either `-R nue' or `-R nue -Q --nue_max_num_vls <int>' to activate Nue.
Note: if `--nue_max_num_vls' is specified and unequal to 1, then QoS support must be turned on, so that SL2VL mappings are valid and applications comply with suggested SLs to avoid credit-loops. For more details on QoS and Nue see below.
The implementation of Nue routing for OpenSM is a 100%-applicable,
balanced, and deadlock-free unicast routing engine (which also configures
multicast tables, see 'Note on multicast' below). The key points of this
algorithm are the following:
- 100% fault-tolerant, oblivious routing strategy
- topology-agnostic, i.e., applicable to every topology (no matter if
topology
is regular, irregular after faults, or random)
- 100% deadlock-free routing within the resource limits (i.e., it never
exceeds the given number of available virtual lanes, and it does not
necessarily require virtual lanes) for every topology
- very good path balancing and therefore high throughput (even better when
using METIS, see notes below)
- QoS (via SLs/VLs) + deadlock-freedom can be combined (since both rely on
VLs), e.g., using VL0-3 for Nue's deadlock-freedom (and 1. QoS level) and
VL4-7 as second QoS level
- forwarding tables are fast to calculate: O(n^2 * log n), however slightly
slower compared to topology-aware routings (for obvious reasons), and
- the path-to-VL mapping only depends on the destination, which may be useful
for scalable, efficient path resolution and caching mechanisms.
From a very high level perspective, Nue routing is similar to DFSSSP (see
above) in the sense that both use Dijkstra and edge weight updates for path
balancing, and paths are mapped to virtual layers assuming a 1:1 mapping of
SL2VL tables. However, the fundamental difference is that Nue routing
doesn't perform the path calculation on the graph representing the real
fabric, and instead routes directly within the channel dependency graph.
This approach allows Nue routing to place routing restrictions (to avoid any
credit-loops) in an on-demand manner, which overcomes the problem of all
other good VL-based algorithms. Meaning, the competitors cannot control or
limit the use of VLs, and might run out of them and have to give up. On the
flip side, Nue may have to use detours for a few routes, and hence cannot
really be considered "shortest-path" routing, because it is
impossible to accomplish deadlock-free, shortest-path routing with an
limited number of available virtual lanes for arbitrary network
topologies.
Note on the use of METIS library with Nue:
Nue routing may has to separate the LIDs into multiple subsets, one for every
virtual layer, if multiple layers are used. Nue has two options to perform
this partitioning (not to be confused with IB partitions); the first is a
fairly simple semi-random assignment of LIDs to layers/subsets, and the
second partitioning uses the METIS library to partition the network graph
into k approximately equal sized parts. The latter approach has shown better
results in terms of path balancing and avoidance of using fallback paths,
and hence it is HIGHLY advised to install/use the METIS library with OpenSM
(enforced via `--enable-metis' configure flag when building OpenSM). For the
rare case, that METIS isn't packaged with the Linux distro, here is a link
to the official website to download and install METIS 5.1.0 manually:
http://glaros.dtc.umn.edu/gkhome/metis/metis/overview
OpenSM's configure script also provides options in case METIS header and
library aren't found in the default path.
Runtime options for Nue:
The behavior of Nue routing can be directly influenced by the osm.conf
parameter (which is also available as command line option):
- nue_max_num_vls: controls/limits the number of virtual lanes/layers which
Nue is allowed to use (detailed explanation in osm.conf file).
Furthermore, Nue supports TRUE and FALSE settings of avoid_throttled_links,
use_ucast_cache, and qos (more on this hereafter); and lmc > 0.
Notes on Quality of Service (QoS):
The advantage of Nue is that it works with AND without QoS being enabled,
i.e., the usage of SLs/VLs for deadlock-freedom can be avoided. Here are the
three possible usage scenarios:
- neither setting `--nue_max_num_vls <int>' nor `-Q': Nue assumes that
only 1
virtual layer (identical to physical network; or OperVLs equal to VL0) is
usable and all paths are to be calculated within this one layer. Hence,
there is no need for special SL2VL mappings in the network and the use of
specific SLs by applications.
- setting `-Q' but not `--nue_max_num_vls <int>': This combination
works like
the previous one, meaning the SL returned for path record requests is not
defined by Nue, since all paths are deadlock-free without using VLs.
However, any separate QoS settings may influence the SL returned to
applications.
- setting `-Q --nue_max_num_vls <int>' with int != 1: In this
configuration,
applications have to query and obey the SL for path records as returned
by Nue because otherwise the deadlock-freedom cannot be guaranteed
anymore. Furthermore, errors in the fabric may require applications to
repath to avoid message deadlocks. Since Nue operates on virtual layer,
admins should configure the SL2VL mapping tables in an homogeneous 1:1
manner across the entire subnet to separate the layers.
As an additional note, using more VLs for Nue usually improves the overall
network throughput, so there are trade offs admins may have to consider when
configuring the subnet manager with Nue routing.
Note on multicast:
The Nue routing engine configures multicast forwarding tables by utilizing a
spanning tree calculation routed at a subnet switch suggested by OpenSM.
This spanning tree for a mcast group will try to use the least overloaded
links (w.r.t the ucast paths-per-link metric/weight) in the fabric. However,
Nue routing currently does not guarantee deadlock-freedom for the set of
multicast routes on all topologies, nor for the combination of deadlock-free
unicast routes with additional multicast routes. Assuming, for a given
topology the calculated mcast routes are dl-free, then an admin may fix the
latter problem by separating the VLs, e.g., using VL0-6 for unicast routing
by specifying `--nue_max_num_vls 7' and utilizing VL7 for multicast.
Routing References
To learn more about deadlock-free routing, see the article "Deadlock Free Message Routing in Multiprocessor Interconnection Networks" by William J Dally and Charles L Seitz (1985).
To learn more about the up/down algorithm, see the article "Effective Strategy to Compute Forwarding Tables for InfiniBand Networks" by Jose Carlos Sancho, Antonio Robles, and Jose Duato at the Universidad Politecnica de Valencia.
To learn more about LASH and the flexibility behind it, the requirement for layers, performance comparisons to other algorithms, see the following articles:
"Layered Routing in Irregular Networks", Lysne et al, IEEE Transactions on Parallel and Distributed Systems, VOL.16, No12, December 2005.
"Routing for the ASI Fabric Manager", Solheim et al. IEEE Communications Magazine, Vol.44, No.7, July 2006.
"Layered Shortest Path (LASH) Routing in Irregular System Area Networks", Skeie et al. IEEE Computer Society Communication Architecture for Clusters 2002.
To learn more about the DFSSSP and SSSP routing algorithm, see the
articles:
J. Domke, T. Hoefler and W. Nagel: Deadlock-Free Oblivious Routing for
Arbitrary Topologies, In Proceedings of the 25th IEEE International Parallel
& Distributed Processing Symposium (IPDPS 2011)
T. Hoefler, T. Schneider and A. Lumsdaine: Optimized Routing for Large-Scale
InfiniBand Networks, In 17th Annual IEEE Symposium on High Performance
Interconnects (HOTI 2009)
To learn more about the Nue routing algorithm, see the article "Routing on the Dependency Graph: A New Approach to Deadlock-Free High-Performance Routing" by J. Domke, T. Hoefler and S. Matsuoka (published in HPDC'16).
Modular Routing Engine
Modular routing engine structure allows for the ease of "plugging" new routing modules.
Currently, only unicast callbacks are supported. Multicast can be added later.
One existing routing module is up-down "updn", which may be activated with '-R updn' option (instead of old '-u').
General usage is: $ opensm -R 'module-name'
There is also a trivial routing module which is able to load LFT tables from a file.
Main features:
- this will load switch LFTs and/or LID matrices (min hops tables)
- this will load switch LFTs according to the path entries introduced
in the file
- no additional checks will be performed (such as "is port
connected",
etc.)
- in case when fabric LIDs were changed this will try to reconstruct
LFTs correctly if endport GUIDs are represented in the file
(in order to disable this, GUIDs may be removed from the file
or zeroed)
The file format is compatible with output of 'ibroute' util and for whole fabric can be generated with dump_lfts.sh script.
To activate file based routing module, use:
opensm -R file -U /path/to/lfts_file
If the lfts_file is not found or is in error, the default routing algorithm is utilized.
The ability to dump switch lid matrices (aka min hops tables) to file and later to load these is also supported.
The usage is similar to unicast forwarding tables loading from a lfts file (introduced by 'file' routing engine), but new lid matrix file name should be specified by -M or --lid_matrix_file option. For example:
opensm -R file -M ./opensm-lid-matrix.dump
The dump file is named ´opensm-lid-matrix.dump´ and will be generated in standard opensm dump directory (/var/log by default) when OSM_LOG_ROUTING logging flag is set.
When routing engine 'file' is activated, but the lfts file is not specified or not cannot be open default lid matrix algorithm will be used.
There is also a switch forwarding tables dumper which generates a file compatible with dump_lfts.sh output. This file can be used as input for forwarding tables loading by 'file' routing engine. Both or one of options -U and -M can be specified together with ´-R file´.
To enable per module logging, configure per_module_logging_file to the per module logging config file name in the opensm options file. To disable, configure per_module_logging_file to (null) there.
The per module logging config file format is a set of lines with module name and logging level as follows:
<module name><separator><logging level>
<module name> is the file name including .c
<separator> is either = , space, or tab
<logging level> is the same levels as used in the coarse/overall
logging as follows:
BIT LOG LEVEL ENABLED
---- -----------------
0x01 - ERROR (error messages)
0x02 - INFO (basic messages, low volume)
0x04 - VERBOSE (interesting stuff, moderate volume)
0x08 - DEBUG (diagnostic, high volume)
0x10 - FUNCS (function entry/exit, very high volume)
0x20 - FRAMES (dumps all SMP and GMP frames)
0x40 - ROUTING (dump FDB routing information)
0x80 - SYS (syslog at LOG_INFO level in addition to OpenSM logging)
Sept 15, 2014 | OpenIB |