Previously we used a fairly simple algorithm in
mca_btl_tcp_proc_insert() to pair local and remote modules. This was a
point in time solution rather than a global optimization problem (where
global means all modules between two peers). The selection logic would
often fail due to pairing interfaces that are not routable for traffic.
The complexity of the selection logic was Θ(n^n), which was expensive.
Due to poor scalability, this logic was only used when the number of
interfaces was less than MAX_PERMUTATION_INTERFACES (default 8). More
details can be found in this ticket:
https://svn.open-mpi.org/trac/ompi/ticket/2031 (The complexity estimates
in the ticket do not match what I calculated from the function)
As a fallback, when interfaces surpassed this threshold, a brute force
O(n^2) double for loop was used to match interfaces.
This commit solves two problems. First, the point-in-time solution is
turned into a global optimization solution. Second, the reachability
framework was used to create a more realistic reachability map. We
switched from using IP/netmask to using the reachability framework,
which supports route lookup. This will help many corner cases as well as
utilize any future development of the reachability framework.
The solution implemented in this commit has a complexity mainly derived
from the bipartite assignment solver. If the local and remote peer both
have the same number of interfaces (n), the complexity of matching will
be O(n^5).
With the decrease in complexity to O(n^5), I calculated and tested
that initialization costs would be 5000 microseconds with 30 interfaces
per node (Likely close to the maximum realistic number of interfaces we
will encounter). For additional datapoints, data up to 300 (a very
unrealistic number) of interfaces was simulated. Up until 150
interfaces, the matching costs will be less than 1 second, climbing to
10 seconds with 300 interfaces. Reflecting on these results, I removed
the suboptimal O(n^2) fallback logic, as it no longer seems necessary.
Data was gathered comparing the scaling of initialization costs with
ranks. For low number of interfaces, the impact of initialization is
negligible. At an interface count of 7-8, the new code has slightly
faster initialization costs. At an interface count of 15, the new code
has slower initialization costs. However, all initialization costs
scale linearly with the number of ranks.
In order to use the reachable function, we populate local and remote
lists of interfaces. We then convert the interface matching problem
into a graph problem. We create a bipartite graph with the local and
remote interfaces as vertices and use negative reachability weights as
costs. Using the bipartite assignment solver, we generate the matches
for the graph. To ensure that both the local and remote process have
the same output, we ensure we mirror their respective inputs for the
graphs. Finally, we store the endpoint matches that we created earlier
in a hash table. This is stored with the btl_index as the key and a
struct mca_btl_tcp_addr_t* as the value. This is then retrieved during
insertion time to set the endpoint address.
Signed-off-by: William Zhang <wilzhang@amazon.com>
Simplify selection of the address to publish for a given BTL TCP
module in the module exchange code. Rather than looping through
all IP addresses associated with a node, looking for one that
matches the kindex of a module, loop over the modules and
use the address stored in the module structure. This also
happens to be the address that the source will use to bind()
in a connect() call, so this should eliminate any confusion
(read: bugs) when an interface has multiple IPs associated with
it.
Refs #5818
Signed-off-by: Brian Barrett <bbarrett@amazon.com>
WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL
All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic.
This commit was SVN r32317.
