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>
This PR removes the constant defining the max attachment address and
replaces it with the largest address that shows up in /proc/self/maps.
This should address issues found on AARCH64 where the max address
may differ based on the configuration.
Since the calculated max address may differ between processes the
max address is sent as part of the modex and stored in the endpoint
data.
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
Make sure to get an RDM provider that can provide both local and
remote communication. We need this check because some providers could
be selected via RXD or RXM, but can't provide local communication, for
example.
Add OPAL_CHECK_OFI_VERSION_GE() m4 macro to check that the Libfabric
we're building against is >= a target version. Use this check in two
places:
1. MTL/OFI: Make sure it is >= v1.5, because the FI_LOCAL_COMM /
FI_REMOTE_COMM constants were introduced in Libfabric API v1.5.
2. BTL/usnic: It already had similar configury to check for Libfabric
>= v1.1, but the usnic component was checking for >= v1.3. So
update the btl/usnic configury to use the new macro and check for
>= v1.3.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This commit fixes an issue discovered in the XPMEM registration cache. It
was possible for a registration to be invalidated by multiple threads
leading to a double-free situation or re-use of an invalidated registration.
This commit fixes the issue by setting the INVALID flag on a registation
when it will be deleted. The flag is set while iterating over the tree
to take advantage of the fact that a registration can not be removed
from the VMA tree by a thread while another thread is traversing the VMA
tree.
References #6524
References #7030Closes#6534
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
related to #7128
The UCX crew is no longer guaranteeing that the UCT API is going to be frozen,
so this is kind of a whack-a-mole problem trying to keep the BTL UCT working
with various changing UCT APIs.
Signed-off-by: Howard Pritchard <howardp@lanl.gov>
This commit fixes a configure bug that caused flow control to be
disabled regardless of the configure options used.
Signed-off-by: Todd Kordenbrock <thkgcode@gmail.com>
Move the prefix area from the head to the body in relevant size
computations. This fixes a problem in high traffic situations where
usNIC may have sent from unregistered memory.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
New MCA param: btl_usnic_max_resends_per_iteration. This is the max
number of resends we'll do in a single pass through usNIC component
progress. This prevents progress from getting stuck in an endless
loop of retransmissions (i.e., if more retransmissions are triggered
during the sending of retransmissions). Specifically: we need to
leave the resend loop to allow receives to happen (which may ACK
messages we have sent previously, and therefore cause pending resends
to be moot).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Significantly increase the default retrans timeout. If the
retrans timeout is too soon, we can end up in a retransmission storm
where the logic will continually re-transmit the same frames during a
single run through the usNIC progress function (because the timer for
a single frame expires before we have run through re-transmitting all
the frames pending re-transmission).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
New MCA parameter: btl_usnic_ack_iteration_delay. Set this to the
number of times through the usNIC component progress function before
sending a standalone ACK (vs. piggy-backing the ACK on any other send
going to the target peer).
Use "ticks" language to clarify that we're really counting the number
of times through the usNIC component DATA_CHANNEL completion check (to
check for incoming messages) -- it has no relation to wall clock time
whatsoever.
Also slightly change the channel-checking scheme in usNIC component
progress: only check the PRIORITY channel once (vs. checking it once,
not finding anything, and then falling through the progress_2() where we
check PRIORITY again and then check the DATA channel).
As before, if our "progress" libevent fires, increment the tick
counter enough to guarantee that all endpoints that need an ACK will
get triggered to send standalone ACKs the next time through progress,
if necessary.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Rename "get_nsec()" to "get_ticks()" to more accurately reflect that
this function has no correlation to wall clock time at all.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
Might as well save a few bytes when sending this struct across the
network via the __opal_attribute_packed__ attribute.
That being said, also re-order the elements in this struct so that
there's no holes to begin with. Do this so that the compiler/runtime
won't effect (slow) unaligned reads/writes because of the
__opal_attribute_packed__ attribute.
The "packed" attribute is really more about defensive programming
(e.g., if we make a mistake and have a hole, "packed" will remove it
for us).
*** Do not bring this commit back to existing/already-released release
branches: it will cause incompatibility, since it effectively changes
the usNIC BTL wire protocol.
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This commit fixes a crash that can occur if a transport
is usable but doesn't have zero-copy support. In this
case do not attempt to use zero-copy and set the max
send size off the bcopy limit.
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
OpenUCX broke the UCT API again in v1.8. This commit updates
btl/uct to fix compilation with current OpenUCX master
(future v1.8). Further changes will likely be needed for
the final release.
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
This commit changes how the single-copy emulation in the vader btl
operates. Before this change the BTL set its put and get limits
based on the max send size. After this change the limits are unset
and the put or get operation is fragmented internally.
References #6568
Signed-off-by: Nathan Hjelm <hjelmn@google.com>
Trying out to run processes via mpirun in Podman containers has shown
that the CMA btl_vader_single_copy_mechanism does not work when user
namespaces are involved.
Creating containers with Podman requires at least user namespaces to be
able to do unprivileged mounts in a container
Even if running the container with user namespace user ID mappings which
result in the same user ID on the inside and outside of all involved
containers, the check in the kernel to allow ptrace (and thus
process_vm_{read,write}v()), fails if the same IDs are not in the same
user namespace.
One workaround is to specify '--mca btl_vader_single_copy_mechanism none'
and this commit adds code to automatically skip CMA if user namespaces
are detected and fall back to MCA_BTL_VADER_EMUL.
Signed-off-by: Adrian Reber <areber@redhat.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>
Remove compatibility code for multiple versions of BTL_IN_OPAL,
BTL_VERSION, and RCACHE_VERSION. This stuff was really only necessary
when we were actively swapping code between multiple release branches
that had large variations in core OMPI infrastructure. These large
variations have now been around for quite a while, so the need for
this "compat" layer is significantly reduced. It hasn't been removed
simply because a few of the "compat" names a slightly more friendly
than the real names (e.g., the SEND/RECV/PUT names).
Signed-off-by: Jeff Squyres <jsquyres@cisco.com>
This commit fixes an error in the 32-bit compare-and-swap atomic support
for Aries networks. The code was incorrectly using the non-fetching
version of cswap which was causing the routing to return
OPAL_ERR_BAD_ARG.
Signed-off-by: Nathan Hjelm <hjelmn@cs.unm.edu>
Avoid printing an error message about ENOTCONN return codes from
getpeername() when handling an incoming connection request. At
this point in the receive state machine, the remote process has
been verified to be a valid OMPI instance. In all-to-all startup
at 4k rank scale, we're seeing this error message when the remote
side drops the connection because it realizes it's the "loser"
in the connection race. We were already doing all the right things,
other than printing a scary error message. So skip the error
message and call it good.
Signed-off-by: Brian Barrett <bbarrett@amazon.com>
free the component mpool in mca_btl_vader_component_close()
and after freeing soem objects that depend on it such as
mca_btl_vader_component.vader_frags_user
Thanks Christoph Niethammer for reporting this.
Refs. open-mpi/ompi#6524
Signed-off-by: Gilles Gouaillardet <gilles@rist.or.jp>
