We currently save the hostname of a proc when we create the ompi_proc_t for it. This was originally done because the only method we had for discovering the host of a proc was to include that info in the modex, and we had to therefore store it somewhere proc-local. Obviously, this ccarried a memory penalty for storing all those strings, and so we added a "cutoff" parameter so that we wouldn't collect hostnames above a certain number of procs.
Unfortunately, this still results in an 8-byte/proc memory cost as we have a char* pointer in the opal_proc_t that is contained in the ompi_proc_t so that we can store the hostname of the other procs if we fall below the cutoff. At scale, this can consume a fair amount of memory.
With the switch to relying on PMIx, there is no longer a need to cache the proc hostnames. Using the "optional" feature of PMIx_Get, we restrict the retrieval to be purely proc-local - i.e., we retrieve the info either via shared memory or from within the proc-internal hash storage (depending upon the active PMIx components). Thus, the retrieval of a hostname is purely a local operation involving no communication.
All RM's are required to provide a complete hostname map of all procs at startup. Thus, we have full access to all hostnames without including them in a modex or having to cache them on each proc. This allows us to remove the char* pointer from the opal_proc_t, saving us 8-bytes/proc.
Unfortunately, PMIx_Get does not currently support the return of a static pointer to memory. Thus, even though PMIx has the hostname in its memory, it can only return a malloc'd version of it. I have therefore ensured that the return from opal_get_proc_hostname is consistently malloc'd and free'd wherever used. This shouldn't be a burden as the hostname is only used in one of two circumstances:
(a) in an error message
(b) in a verbose output for debugging purposes
Thus, there should be no performance penalty associated with the malloc/free requirement. PMIx will eventually be returning static pointers, and so we can eventually simplify this method and return a "const char*" - but as noted, this really isn't an issue even today.
Signed-off-by: Ralph Castain <rhc@pmix.org>
Will be replaced by PRRTE. Ensure that OMPI and OPAL layers build
without reference to ORTE. Setup opal/pmix framework to be static.
Remove support for all PMI-1 and PMI-2 libraries. Add support for
"external" pmix component as well as internal v4 one.
remove orte: misc fixes
- UCX fixes
- VPATH issue
- oshmem fixes
- remove useless definition
- Add PRRTE submodule
- Get autogen.pl to traverse PRRTE submodule
- Remove stale orcm reference
- Configure embedded PRRTE
- Correctly pass the prefix to PRRTE
- Correctly set the OMPI_WANT_PRRTE am_conditional
- Move prrte configuration to the end of OMPI's configure.ac
- Make mpirun a symlink to prun, when available
- Fix makedist with --no-orte/--no-prrte option
- Add a `--no-prrte` option which is the same as the legacy
`--no-orte` option.
- Remove embedded PMIx tarball. Replace it with new submodule
pointing to OpenPMIx master repo's master branch
- Some cleanup in PRRTE integration and add config summary entry
- Correctly set the hostname
- Fix locality
- Fix singleton operations
- Fix support for "tune" and "am" options
Signed-off-by: Ralph Castain <rhc@pmix.org>
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
Signed-off-by: Joshua Hursey <jhursey@us.ibm.com>
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>
Due to IF_NAMESIZE being a reused and conditionally defined macro,
issues could arise from macro mismatches. In particular, in cases where
opal/util/if.h is included, but net/if.h is not, IF_NAMESIZE will be 32.
If net/if.h is included on Linux systems, IF_NAMESIZE will be 16. This
can cause a mismatch when using the same macro on a system. Thus
different parts of the code can have differring ideas on the size of a
structure containing a char name[IF_NAMESIZE]. To avoid this error case,
we avoid reusing the IF_NAMESIZE macro and instead define our own as
OPAL_IF_NAMESIZE.
Signed-off-by: William Zhang <wilzhang@amazon.com>
After the OPAL_MODEX_RECV call, remote_addrs was not freed in the error
path. Moved the free call into cleanup to ensure we always free this
memory before leaving the function.
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>
While trying to debug #3035, it's not clear whether there is
an issue with the modex data or printing the address list.
Print the number of endpoints on the error, which will help
determine which case is happening to Cisco.
Signed-off-by: Brian Barrett <bbarrett@amazon.com>
The Open MPI code base assumed that asprintf always behaved like
the FreeBSD variant, where ptr is set to NULL on error. However,
the C standard (and Linux) only guarantee that the return code will
be -1 on error and leave ptr undefined. Rather than fix all the
usage in the code, we use opal_asprintf() wrapper instead, which
guarantees the BSD-like behavior of ptr always being set to NULL.
In addition to being correct, this will fix many, many warnings
in the Open MPI code base.
Signed-off-by: Brian Barrett <bbarrett@amazon.com>
Per
https://github.com/open-mpi/ompi/issues/3035#issuecomment-426085673,
it looks like the IP address for a given interface is being stashed in
two places: on the endpoint and on the module.
1. On the endpoint, it is storing the moral equivalent of a
(struct sockaddr_in.sin_addr).
2. On the module, it is storing a full (struct sockaddr_storage).
The call to opal_net_get_hostname() expects a full (struct sockaddr*)
-- not just the stripped-down (struct sockaddr_in.sin_addr). Hence,
when the original code was passing in the endpoint's (struct
sockaddr_in.sin_addr) and opal_net_get_hostname() was treating it
like a (struct sockaddr), hilarity ensued (i.e., we got the wrong
output).
This commit eliminates the call to opal_net_get_hostname() and just
calls inet_ntop() directly to convert the (struct
sockaddr_in.sin_addr) to a string.
NOTE: Per the github comment cited above, there can be a disparity
between the IP address cached on the endpoint vs. the IP address
cached on the module. This only happens with interfaces that have
more than one IP address. This commit does not fix that issue.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
The current cast is *functional*, but isn't really the way it should
be done. This commit makes the cast the way it should be done.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Fix two facepalms:
1. The "uint32" in the hash map functions refer to the *key* size, not
the *value* size. The values are always 64 bits.
2. Pass the straight value to the "set" functions -- not the pointer
to the value.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
The giant size of the TCP proc struct is causing a problem in some
environments (because it is allocated on the stack), and it was too
big, anyway.
Instead, use a hash map. That way, it starts small and can grow if it
needs to. It also makes no assumptions about the values of the kernel
interface indexes.
Fixes#5292.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Fix case where the btl_tcp_links MCA parameter is used to create multiple TCP connections between peers.
Three issues were resulting in hangs during large message transfer:
* The 2nd..btl_tcp_link connections were dropped during establishment because the per-process
address check was binary, rather than a count
* The accept handler would not skip a btl module that was already in use, resulting in all
connections for a given address being vectored to a single btl
* Multiple addresses in the same subnet caused connections to be
stalled, as the receiver would always use the same (first) address
found. Binding the outgoing connection solves this issue
* Lastly fix race condition created by connections being started at the exact same time
by accpeting connections not in the closed state, allowing endpoint_accept to resolve
dispute
Signed-off-by: Jordan Cherry <cherryj@amazon.com>
Their is racing condition in TCP connection establishment
during simultaneous handshake. This PR handles the fix for
it.
Signed-off-by: Mohan Gandhi <mohgan@amazon.com>
This is mostly for error cases, where we need to release the
newly created proc. Currently the code deadlocks because the endpoint
lock is help at the release and the lock is not recursive.
Aslo added some code to print the IP addresses that don't match during
the TCP connection step.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
Add a verbose to show all the failed attempts to match the
remote interfaces based on the modex info.
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
This patch is based on the "RFC: Reenabling the TCP BTL over local
interfaces (when specifically requested)". It removes the hardcoded
exception for the local devices that has been enforced by the
TCP BTL. Instead, we exclude the local interface only via the
exclude MCA (both IPv4 and IPv6 local addresses are already in the
default if_exclude), which is also the behavior currently described in
our README file.
This commit fixes a race between a thread calling the tcp btl's
add_procs and a thread processing an incomming connection. The race
occured because the add_procs thread adds a newly created proc object
to the hash table *before* the object is fully initialized. The
connection thread then attempts to use the object before the endpoints
array on the object has beeen allocation. The fix is to only add the
proc to the hash table after it has been completely initialized.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
This commit makes two changes to the tcp btl:
- If a tcp proc does not exist when handling a new connection create
a new proc and use it. The current implementation uses the
opal_proc_by_name() function to get the opal_proc_t then calls
add_procs on all btl modules. It may be sufficient to just call
add_procs until an endpoint is created so this may change somewhat.
- In add_procs add a check for an existing endpoint before creating
one.
Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Bring Slurm PMI-1 component online
Bring the s2 component online
Little cleanup - let the various PMIx modules set the process name during init, and then just raise it up to the ORTE level. Required as the different PMI environments all pass the jobid in different ways.
Bring the OMPI pubsub/pmi component online
Get comm_spawn working again
Ensure we always provide a cpuset, even if it is NULL
pmix/cray: adjust cray pmix component for pmix
Make changes so cray pmix can work within the integrated
ompi/pmix framework.
Bring singletons back online. Implement the comm_spawn operation using pmix - not tested yet
Cleanup comm_spawn - procs now starting, error in connect_accept
Complete integration
Also add another (superflous but symmetric) continue statement.
This missing "continue" statement allows IPv4 "private network"
matches to fall through and allow IPv6 matches to be made -- thereby
overriding the IPv4 match that was already made.
Fixes#585 (although several of the other issues identified on #585
still exist, the primary / initial bug that was reported there is now
fixed).
while cleaning up after receiving a zero byte on the connect socket
(localyy started connection), while another was trying to accept a
new connection from the same peer. Create a zero-timed event and
delocalize the accept into a timer_event.
Add support for registering an error callback, that can be used when a
connection is discovered as failed during the initialization process.
inserted in the ompi_proc_list as soon as it is created and it
is removed only upon the call to the destructor. In ompi_proc_finalize
we loop over all procs in ompi_proc_finalize and release them once.
However, as a proc is not removed from this list right away, we
decrease the ref count for each proc until it reach zero and the
proc is finally removed. Thus, we cannot clean the BML/BTL after
the call the ompi_proc_finalize.
A quick fix is to delay the call to ompi_proc_finalize until all
other frameworks have been finalized, and then the behavior
depicted above will give the expected outcome.
We recognize that this means other users of OPAL will need to "wrap" the opal_process_name_t if they desire to abstract it in some fashion. This is regrettable, and we are looking at possible alternatives that might mitigate that requirement. Meantime, however, we have to put the needs of the OMPI community first, and are taking this step to restore hetero and SPARC support.
Properly setup the opal_process_info structure early in the initialization procedure. Define the local hostname right at the beginning of opal_init so all parts of opal can use it. Overlay that during orte_init as the user may choose to remove fqdn and strip prefixes during that time. Setup the job_session_dir and other such info immediately when it becomes available during orte_init.
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
