WIP: Backports for 0.6.2 #24519
WIP: Backports for 0.6.2 #24519
Conversation
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
c75018a to
eaf7368
Compare
|
Looks like the builds are freezing, not sure why yet. |
|
The following testsets does not have any output on any of the Travis bots: |
|
Has 943c8e5 been considered for the list? (Perhaps 6.3?) This commit fixes a issue that requires some unfortunate work-arounds in JuliaCall and rjulia. |
|
@phaverty, according to Yichao that commit is irrelevant. Did you mean a different one? |
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
eaf7368 to
c767cdc
Compare
|
Looks like |
ararslan
force-pushed
the
aa/backports-0.6.2
branch
2 times, most recently
from
3b9ab8d to
0bcb332
Compare
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
0bcb332 to
b8bb2b0
Compare
|
Now all tests are passing on AppVeyor and Travis OS X, but the socket tests are failing on Travis Linux. Oddly enough, the socket tests pass on Nanosoldier (x86-64 Linux). |
|
Could #24503 be included as it fixes a severe regression on windows? |
|
Yes it will be. |
|
Does anyone have any ideas about the socket tests on Travis Linux? |
|
Seems like it should be easy to pull in #24540. |
|
That one isn't relevant for backporting since FFTW isn't deprecated on 0.6. |
|
There are a couple of commits marked for backport that I've requested the original authors backport since I wasn't able to determine how to do it correctly, but the current state of the PR contains nearly all of the relevant commits for 0.6.2. |
|
#24530 would be good to have in my opinion, I think it should be good to merge. |
|
Sure. Just needs to be merged and I can backport it. 😉 |
|
Waiting on a couple more commits but might as well check performance in the meantime. @nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan |
|
Hmmm. @nanosoldier |
|
The SIGABRT on Travis Linux when building the documentation is very concerning but I can't reproduce it. |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan |
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
d9be6af to
29b6dad
Compare
|
Should #24112 be included? |
|
The latest state of this branch now results in additional test failures for CuArrays, CUFFT and CUBLAS. The last two seem slightly unreliable, so @MikeInnes could you look into the CuArrays test failure? |
|
@amitmurthy Should I revert #21818 in 0.6.2 or is there a better solution? |
|
Here it is diff --git a/NEWS.md b/NEWS.md
index 62e2983c8e..13d26cdb4d 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -259,7 +259,8 @@ This section lists changes that do not have deprecation warnings.
rather than from environment variables ([#19636]).
* Workers now listen on an ephemeral port assigned by the OS. Previously workers would
- listen on the first free port available from 9009 ([#21818]).
+ listen on the first free port available from 9009 ([#21818]). Version 0.6.1 only.
+ Reverted in 0.6.2
Library improvements
diff --git a/base/distributed/cluster.jl b/base/distributed/cluster.jl
index c12f8ed847..1d1f74d7f2 100644
--- a/base/distributed/cluster.jl
+++ b/base/distributed/cluster.jl
@@ -153,7 +153,7 @@ function start_worker(out::IO, cookie::AbstractString)
init_worker(cookie)
interface = IPv4(LPROC.bind_addr)
if LPROC.bind_port == 0
- (port, sock) = listenany(interface, UInt16(0))
+ (port, sock) = listenany(interface, UInt16(9009))
LPROC.bind_port = port
else
sock = listen(interface, LPROC.bind_port)
diff --git a/doc/src/manual/parallel-computing.md b/doc/src/manual/parallel-computing.md
index 8b3c15dc37..4d2fb1e97d 100644
--- a/doc/src/manual/parallel-computing.md
+++ b/doc/src/manual/parallel-computing.md
@@ -1231,8 +1231,8 @@ as local laptops, departmental clusters, or even the cloud. This section covers
requirements for the inbuilt `LocalManager` and `SSHManager`:
* The master process does not listen on any port. It only connects out to the workers.
- * Each worker binds to only one of the local interfaces and listens on an ephemeral port number
- assigned by the OS.
+ * Each worker binds to only one of the local interfaces and listens on the first free port starting
+ from `9009`.
* `LocalManager`, used by `addprocs(N)`, by default binds only to the loopback interface. This means
that workers started later on remote hosts (or by anyone with malicious intentions) are unable
to connect to the cluster. An `addprocs(4)` followed by an `addprocs(["remote_host"])` will fail.
@@ -1250,9 +1250,8 @@ requirements for the inbuilt `LocalManager` and `SSHManager`:
authenticated via public key infrastructure (PKI). Authentication credentials can be supplied
via `sshflags`, for example ```sshflags=`-e <keyfile>` ```.
- In an all-to-all topology (the default), all workers connect to each other via plain TCP sockets.
- The security policy on the cluster nodes must thus ensure free connectivity between workers for
- the ephemeral port range (varies by OS).
+ Note that worker-worker connections are still plain TCP and the local security policy on the remote
+ cluster must allow for free connections between worker nodes, at least for ports 9009 and above.
Securing and encrypting all worker-worker traffic (via SSH) or encrypting individual messages
can be done via a custom ClusterManager. |
|
Great, thanks @amitmurthy! @MikeInnes, any word on CuArrays? I don't have the right system to be able to test myself so I'll need some help ensuring that this PR isn't breaking for the GPU ecosystem. |
|
@amitmurthy It looks like the tests may also have to be adjusted to accommodate that change? |
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
86a0e9c to
0ad7336
Compare
|
I think I figured out how to fix the test: revert the changes made to test/socket.jl in #21818. |
|
Hmm, no that should be a problem. What is the error you are seeing? |
|
I think it timed out listening for a connection on port 0 rather than 9009. It was in the previous Travis macOS log, which has since been overwritten. |
|
Looks like my fix did not fix it: https://travis-ci.org/JuliaLang/julia/jobs/306991362#L5678 |
|
The socket tests need to be left as is. It tests the bugfix in #21818 for listenany() with port 0. Let me build and test this branch locally. Will take a while to build all dependencies. |
ararslan
force-pushed
the
aa/backports-0.6.2
branch
from
0ad7336 to
eda0b7c
Compare
|
Sounds good, thanks. In the meantime I've reverted the change to the tests. |
amitmurthy
force-pushed
the
aa/backports-0.6.2
branch
from
eda0b7c to
ed763c5
Compare
|
Changing Pushed directly to this branch. |
|
I'll also test it out later on anubis. Just realized that Travis Linux
setups do not support client socket reuse.
|
|
CuArrays passes tests for me, although with quite a few warnings about invalid debug info being generated. I don't know if it's just a local issue, but on my usual GPU machine I can't compile this branch (it's asking me what files I want to apply some LLVM patches to). |
That's a change to CUDAnative which'll disappear when the version string matches
Probably a dirty LLVM source tree from moving between eg. |
amitmurthy
force-pushed
the
aa/backports-0.6.2
branch
from
ed763c5 to
6736f45
Compare
|
@ararslan , the run was successful on Anubis. However, taking a conservative approach I reverted to both 1) listening on 9009 onwards and 2) disabling client socket reuse on all platforms. |
|
Might as well @nanosoldier |
|
Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @ararslan |
|
Latest PkgEval results: https://gist.github.com/ararslan/0b1112ea62d9f37ac10aac6c4b624572
Looks like this branch is now non-breaking. |
|
Unless there are any objections or someone beats me to it, I'm going to merge this tomorrow. |
|
That was quite some patient work! |
No description provided.