TORUS-2QOS(8) OpenIB Management TORUS-2QOS(8)NAME
torus-2QoS - Routing engine for OpenSM subnet manager
DESCRIPTION
Torus-2QoS is routing algorithm designed for large-scale 2D/3D torus
fabrics. The torus-2QoS routing engine can provide the following func‐
tionality on a 2D/3D torus:
– Routing that is free of credit loops.
– Two levels of Quality of Service (QoS), assuming switches and chan‐
nel adapters support eight data VLs.
– The ability to route around a single failed switch, and/or multiple
failed links, without
– introducing credit loops, or
– changing path SL values.
– Very short run times, with good scaling properties as fabric size
increases.
UNICAST ROUTING
Unicast routing in torus-2QoS is based on Dimension Order Routing
(DOR). It avoids the deadlocks that would otherwise occur in a DOR-
routed torus using the concept of a dateline for each torus dimension.
It encodes into a path SL which datelines the path crosses, as follows:
sl = 0;
for (d = 0; d < torus_dimensions; d++) {
/* path_crosses_dateline(d) returns 0 or 1 */
sl |= path_crosses_dateline(d) << d;
}
On a 3D torus this consumes three SL bits, leaving one SL bit unused.
Torus-2QoS uses this SL bit to implement two QoS levels.
Torus-2QoS also makes use of the output port dependence of switch SL2VL
maps to encode into one VL bit the information encoded in three SL
bits. It computes in which torus coordinate direction each inter-
switch link "points", and writes SL2VL maps for such ports as follows:
for (sl = 0; sl < 16; sl++) {
/* cdir(port) computes which torus coordinate direction
* a switch port "points" in; returns 0, 1, or 2
*/
sl2vl(iport,oport,sl) = 0x1 & (sl >> cdir(oport));
}
Thus, on a pristine 3D torus, i.e., in the absence of failed fabric
switches, torus-2QoS consumes eight SL values (SL bits 0-2) and two VL
values (VL bit 0) per QoS level to provide deadlock-free routing.
Torus-2QoS routes around link failure by "taking the long way around"
any 1D ring interrupted by link failure. For example, consider the 2D
6x5 torus below, where switches are denoted by [+a-zA-Z]:
| | | | | |
4 --+----+----+----+----+----+--
| | | | | |
3 --+----+----+----D----+----+--
| | | | | |
2 --+----+----I----r----+----+--
| | | | | |
1 --m----S----n----T----o----p--
| | | | | |
y=0 --+----+----+----+----+----+--
| | | | | |
x=0 1 2 3 4 5
For a pristine fabric the path from S to D would be S-n-T-r-D. In the
event that either link S-n or n-T has failed, torus-2QoS would use the
path S-m-p-o-T-r-D. Note that it can do this without changing the path
SL value; once the 1D ring m-S-n-T-o-p-m has been broken by failure,
path segments using it cannot contribute to deadlock, and the x-direc‐
tion dateline (between, say, x=5 and x=0) can be ignored for path seg‐
ments on that ring.
One result of this is that torus-2QoS can route around many simultane‐
ous link failures, as long as no 1D ring is broken into disjoint seg‐
ments. For example, if links n-T and T-o have both failed, that ring
has been broken into two disjoint segments, T and o-p-m-S-n.
Torus-2QoS checks for such issues, reports if they are found, and
refuses to route such fabrics.
Note that in the case where there are multiple parallel links between a
pair of switches, torus-2QoS will allocate routes across such links in
a round-robin fashion, based on ports at the path destination switch
that are active and not used for inter-switch links. Should a link
that is one of several such parallel links fail, routes are redis‐
tributed across the remaining links. When the last of such a set of
parallel links fails, traffic is rerouted as described above.
Handling a failed switch under DOR requires introducing into a path at
least one turn that would be otherwise "illegal", i.e., not allowed by
DOR rules. Torus-2QoS will introduce such a turn as close as possible
to the failed switch in order to route around it.
In the above example, suppose switch T has failed, and consider the
path from S to D. Torus-2QoS will produce the path S-n-I-r-D, rather
than the S-n-T-r-D path for a pristine torus, by introducing an early
turn at n. Normal DOR rules will cause traffic arriving at switch I to
be forwarded to switch r; for traffic arriving from I due to the
"early" turn at n, this will generate an "illegal" turn at I.
Torus-2QoS will also use the input port dependence of SL2VL maps to set
VL bit 1 (which would be otherwise unused) for y-x, z-x, and z-y turns,
i.e., those turns that are illegal under DOR. This causes the first
hop after any such turn to use a separate set of VL values, and pre‐
vents deadlock in the presence of a single failed switch.
For any given path, only the hops after a turn that is illegal under
DOR can contribute to a credit loop that leads to deadlock. So in the
example above with failed switch T, the location of the illegal turn at
I in the path from S to D requires that any credit loop caused by that
turn must encircle the failed switch at T. Thus the second and later
hops after the illegal turn at I (i.e., hop r-D) cannot contribute to a
credit loop because they cannot be used to construct a loop encircling
T. The hop I-r uses a separate VL, so it cannot contribute to a credit
loop encircling T.
Extending this argument shows that in addition to being capable of
routing around a single switch failure without introducing deadlock,
torus-2QoS can also route around multiple failed switches on the condi‐
tion they are adjacent in the last dimension routed by DOR. For exam‐
ple, consider the following case on a 6x6 2D torus:
| | | | | |
5 --+----+----+----+----+----+--
| | | | | |
4 --+----+----+----D----+----+--
| | | | | |
3 --+----+----I----u----+----+--
| | | | | |
2 --+----+----q----R----+----+--
| | | | | |
1 --m----S----n----T----o----p--
| | | | | |
y=0 --+----+----+----+----+----+--
| | | | | |
x=0 1 2 3 4 5
Suppose switches T and R have failed, and consider the path from S to
D. Torus-2QoS will generate the path S-n-q-I-u-D, with an illegal turn
at switch I, and with hop I-u using a VL with bit 1 set.
As a further example, consider a case that torus-2QoS cannot route
without deadlock: two failed switches adjacent in a dimension that is
not the last dimension routed by DOR; here the failed switches are O
and T:
| | | | | |
5 --+----+----+----+----+----+--
| | | | | |
4 --+----+----+----+----+----+--
| | | | | |
3 --+----+----+----+----D----+--
| | | | | |
2 --+----+----I----q----r----+--
| | | | | |
1 --m----S----n----O----T----p--
| | | | | |
y=0 --+----+----+----+----+----+--
| | | | | |
x=0 1 2 3 4 5
In a pristine fabric, torus-2QoS would generate the path from S to D as
S-n-O-T-r-D. With failed switches O and T, torus-2QoS will generate
the path S-n-I-q-r-D, with illegal turn at switch I, and with hop I-q
using a VL with bit 1 set. In contrast to the earlier examples, the
second hop after the illegal turn, q-r, can be used to construct a
credit loop encircling the failed switches.
MULTICAST ROUTING
Since torus-2QoS uses all four available SL bits, and the three data VL
bits that are typically available in current switches, there is no way
to use SL/VL values to separate multicast traffic from unicast traffic.
Thus, torus-2QoS must generate multicast routing such that credit loops
cannot arise from a combination of multicast and unicast path segments.
It turns out that it is possible to construct spanning trees for multi‐
cast routing that have that property. For the 2D 6x5 torus example
above, here is the full-fabric spanning tree that torus-2QoS will con‐
struct, where "x" is the root switch and each "+" is a non-root switch:
4 + + + + + +
| | | | | |
3 + + + + + +
| | | | | |
2 +----+----+----x----+----+
| | | | | |
1 + + + + + +
| | | | | |
y=0 + + + + + +
x=0 1 2 3 4 5
For multicast traffic routed from root to tip, every turn in the above
spanning tree is a legal DOR turn.
For traffic routed from tip to root, and some traffic routed through
the root, turns are not legal DOR turns. However, to construct a
credit loop, the union of multicast routing on this spanning tree with
DOR unicast routing can only provide 3 of the 4 turns needed for the
loop.
In addition, if none of the above spanning tree branches crosses a
dateline used for unicast credit loop avoidance on a torus, and if mul‐
ticast traffic is confined to SL 0 or SL 8 (recall that torus-2QoS uses
SL bit 3 to differentiate QoS level), then multicast traffic also can‐
not contribute to the "ring" credit loops that are otherwise possible
in a torus.
Torus-2QoS uses these ideas to create a master spanning tree. Every
multicast group spanning tree will be constructed as a subset of the
master tree, with the same root as the master tree.
Such multicast group spanning trees will in general not be optimal for
groups which are a subset of the full fabric. However, this compromise
must be made to enable support for two QoS levels on a torus while pre‐
venting credit loops.
In the presence of link or switch failures that result in a fabric for
which torus-2QoS can generate credit-loop-free unicast routes, it is
also possible to generate a master spanning tree for multicast that
retains the required properties. For example, consider that same 2D
6x5 torus, with the link from (2,2) to (3,2) failed. Torus-2QoS will
generate the following master spanning tree:
4 + + + + + +
| | | | | |
3 + + + + + +
| | | | | |
2 --+----+----+ x----+----+--
| | | | | |
1 + + + + + +
| | | | | |
y=0 + + + + + +
x=0 1 2 3 4 5
Two things are notable about this master spanning tree. First, assum‐
ing the x dateline was between x=5 and x=0, this spanning tree has a
branch that crosses the dateline. However, just as for unicast, cross‐
ing a dateline on a 1D ring (here, the ring for y=2) that is broken by
a failure cannot contribute to a torus credit loop.
Second, this spanning tree is no longer optimal even for multicast
groups that encompass the entire fabric. That, unfortunately, is a
compromise that must be made to retain the other desirable properties
of torus-2QoS routing.
In the event that a single switch fails, torus-2QoS will generate a
master spanning tree that has no "extra" turns by appropriately select‐
ing a root switch. In the 2D 6x5 torus example, assume now that the
switch at (3,2), i.e., the root for a pristine fabric, fails.
Torus-2QoS will generate the following master spanning tree for that
case:
|
4 + + + + + +
| | | | | |
3 + + + + + +
| | | | |
2 + + + + +
| | | | |
1 +----+----x----+----+----+
| | | | | |
y=0 + + + + + +
|
x=0 1 2 3 4 5
Assuming the y dateline was between y=4 and y=0, this spanning tree has
a branch that crosses a dateline. However, again this cannot contrib‐
ute to credit loops as it occurs on a 1D ring (the ring for x=3) that
is broken by a failure, as in the above example.
TORUS TOPOLOGY DISCOVERY
The algorithm used by torus-2QoS to construct the torus topology from
the undirected graph representing the fabric requires that the radix of
each dimension be configured via torus-2QoS.conf. It also requires
that the torus topology be "seeded"; for a 3D torus this requires con‐
figuring four switches that define the three coordinate directions of
the torus.
Given this starting information, the algorithm is to examine the cube
formed by the eight switch locations bounded by the corners (x,y,z) and
(x+1,y+1,z+1). Based on switches already placed into the torus topol‐
ogy at some of these locations, the algorithm examines 4-loops of
inter-switch links to find the one that is consistent with a face of
the cube of switch locations, and adds its swiches to the discovered
topology in the correct locations.
Because the algorithm is based on examining the topology of 4-loops of
links, a torus with one or more radix-4 dimensions requires extra ini‐
tial seed configuration. See torus-2QoS.conf(5) for details.
Torus-2QoS will detect and report when it has insufficient configura‐
tion for a torus with radix-4 dimensions.
In the event the torus is significantly degraded, i.e., there are many
missing switches or links, it may happen that torus-2QoS is unable to
place into the torus some switches and/or links that were discovered in
the fabric, and will generate a warning in that case. A similar condi‐
tion occurs if torus-2QoS is misconfigured, i.e., the radix of a torus
dimension as configured does not match the radix of that torus dimen‐
sion as wired, and many switches/links in the fabric will not be placed
into the torus.
QUALITY OF SERVICE CONFIGURATION
OpenSM will not program switchs and channel adapters with SL2VL maps or
VL arbitration configuration unless it is invoked with -Q. Since
torus-2QoS depends on such functionality for correct operation, always
invoke OpenSM with -Q when torus-2QoS is in the list of routing
engines.
Any quality of service configuration method supported by OpenSM will
work with torus-2QoS, subject to the following limitations and consid‐
erations.
For all routing engines supported by OpenSM except torus-2QoS, there is
a one-to-one correspondence between QoS level and SL. Torus-2QoS can
only support two quality of service levels, so only the high-order bit
of any SL value used for unicast QoS configuration will be honored by
torus-2QoS.
For multicast QoS configuration, only SL values 0 and 8 should be used
with torus-2QoS.
Since SL to VL map configuration must be under the complete control of
torus-2QoS, any configuration via qos_sl2vl, qos_swe_sl2vl, etc., must
and will be ignored, and a warning will be generated.
Torus-2QoS uses VL values 0-3 to implement one of its supported QoS
levels, and VL values 4-7 to implement the other. Hard-to-diagnose
application issues may arise if traffic is not delivered fairly across
each of these two VL ranges. Torus-2QoS will detect and warn if VL
arbitration is configured unfairly across VLs in the range 0-3, and
also in the range 4-7. Note that the default OpenSM VL arbitration
configuration does not meet this constraint, so all torus-2QoS users
should configure VL arbitration via qos_vlarb_high, qos_vlarb_low, etc.
OPERATIONAL CONSIDERATIONS
Any routing algorithm for a torus IB fabric must employ path SL values
to avoid credit loops. As a result, all applications run over such
fabrics must perform a path record query to obtain the correct path SL
for connection setup. Applications that use rdma_cm for connection
setup will automatically meet this requirement.
If a change in fabric topology causes changes in path SL values
required to route without credit loops, in general all applications
would need to repath to avoid message deadlock. Since torus-2QoS has
the ability to reroute after a single switch failure without changing
path SL values, repathing by running applications is not required when
the fabric is routed with torus-2QoS.
Torus-2QoS can provide unchanging path SL values in the presence of
subnet manager failover provided that all OpenSM instances have the
same idea of dateline location. See torus-2QoS.conf(5) for details.
Torus-2QoS will detect configurations of failed switches and links that
prevent routing that is free of credit loops, and will log warnings and
refuse to route. If "no_fallback" was configured in the list of OpenSM
routing engines, then no other routing engine will attempt to route the
fabric. In that case all paths that do not transit the failed compo‐
nents will continue to work, and the subset of paths that are still
operational will continue to remain free of credit loops. OpenSM will
continue to attempt to route the fabric after every sweep interval, and
after any change (such as a link up) in the fabric topology. When the
fabric components are repaired, full functionality will be restored.
In the event OpenSM was configured to allow some other engine to route
the fabric if torus-2QoS fails, then credit loops and message deadlock
are likely if torus-2QoS had previously routed the fabric successfully.
Even if the other engine is capable of routing a torus without credit
loops, applications that built connections with path SL values granted
under torus-2QoS will likely experience message deadlock under routing
generated by a different engine, unless they repath.
To verify that a torus fabric is routed free of credit loops, use ibdm‐
chk to analyze data collected via ibdiagnet -vlr.
FILES
/etc/rdma/opensm.conf
default OpenSM config file.
/etc/rdma/qos-policy.conf
default QoS policy config file.
/etc/rdma/torus-2QoS.conf
default torus-2QoS config file.
SEE ALSOopensm(8), torus-2QoS.conf(5), ibdiagnet(1), ibdmchk(1), rdma_cm(7).
OpenIB November 10, 2010 TORUS-2QOS(8)