These are chat archives for elemental/chat

14th
Nov 2016
Ryan H. Lewis
@rhl-
Nov 14 2016 00:00
semialgebraic is fairly generic
I mean, i don’t understand that conjecture
l1 is not differentiable
so strictly speaking BFGS is not well defined.
Jack Poulson
@poulson
Nov 14 2016 00:01
yes, but the claim is that the non-differentiability should not be encountered with probability one if one randomly initializes and uses an inexact line search
Ryan H. Lewis
@rhl-
Nov 14 2016 00:02
so, that can’t be true
Jack Poulson
@poulson
Nov 14 2016 00:02
why not?
Ryan H. Lewis
@rhl-
Nov 14 2016 00:02
well..
i mean, i’m just thinking of a function which has minimizer zero
but it not differentiable at zero
but is as smooth as possible elsewhere.
Jack Poulson
@poulson
Nov 14 2016 00:03
but perhaps one never gets there
only converges to it
Ryan H. Lewis
@rhl-
Nov 14 2016 00:04
to the minimizer?
Jack Poulson
@poulson
Nov 14 2016 00:04
yes
Ryan H. Lewis
@rhl-
Nov 14 2016 00:04
like f(x) = x^2 unless x = 0, in which case it is 7 ?
or something?
Jack Poulson
@poulson
Nov 14 2016 00:04
e.g., for the single-variable l1 penalty |x|, one could have iterates x_n = (-1/2)^n
Ryan H. Lewis
@rhl-
Nov 14 2016 00:04
i see..
Jack Poulson
@poulson
Nov 14 2016 00:04
that isn't locally lipschitz
Ryan H. Lewis
@rhl-
Nov 14 2016 00:05
ok, good point
Jack Poulson
@poulson
Nov 14 2016 00:06
I would recommend some of the experiments in the Lewis/Overton paper as good "practical" BFGS tests
or perhaps one with an l1 penalty
Ryan H. Lewis
@rhl-
Nov 14 2016 00:15
@poulson how does the linode thing work? can I build a fedora image or something as well?
Jack Poulson
@poulson
Nov 14 2016 00:24
yes, I think so
but it might increase the billing (not sure though)
FYI: elemental/Elemental@6fd612e
Ryan H. Lewis
@rhl-
Nov 14 2016 00:26
Jack Poulson
@poulson
Nov 14 2016 00:28
that should be fixed now
Ryan H. Lewis
@rhl-
Nov 14 2016 00:28
@poulson something is very wrong: rhl@ubuntu:~\$ mkdir .ssh
mkdir: cannot create directory ‘.ssh’: No space left on device
Filesystem Size Used Avail Use% Mounted on
/dev/root 48G 47G 0 100% /
devtmpfs 3.9G 0 3.9G 0% /dev
tmpfs 3.9G 0 3.9G 0% /dev/shm
tmpfs 3.9G 9.6M 3.9G 1% /run
tmpfs 5.0M 0 5.0M 0% /run/lock
tmpfs 3.9G 0 3.9G 0% /sys/fs/cgroup
tmpfs 799M 0 799M 0% /run/user/0
tmpfs 799M 0 799M 0% /run/user/1002
Jack Poulson
@poulson
Nov 14 2016 00:30
yea, I'm looking into that, as it doesn't make sense
I haven't really done anything on that machine
Ryan H. Lewis
@rhl-
Nov 14 2016 00:31
if you look at the usage graphs on the web
there was serious usage a few months back in terms of Disk I/O
Jack Poulson
@poulson
Nov 14 2016 00:33
hopefully the machine wasn't hacked
I did get an email about significant usage a while back
Ryan H. Lewis
@rhl-
Nov 14 2016 00:33
and the next month
for not using it, it seems like it was being used quite a bit.
not sure what the units are
maybe its just a scaling issue on these graphs
Jack Poulson
@poulson
Nov 14 2016 00:34
Jenkins builds are being automatically triggered
hopefully they aren't filling up the disk with logs
Ryan H. Lewis
@rhl-
Nov 14 2016 00:34
oh
yeah, im like 100% sure that is the problem
jenkins is like DDOSing the box
Jack Poulson
@poulson
Nov 14 2016 00:35
:-p
/var/lib/jenkins is the problem
Ryan H. Lewis
@rhl-
Nov 14 2016 00:36
maybe your not removing old builds
Jack Poulson
@poulson
Nov 14 2016 00:36
I think not
maybe only the past 10 should be kept
Ryan H. Lewis
@rhl-
Nov 14 2016 00:36
Also, the github integration is not working
yeah, last 10 or 20 is pretty normal.
we use last 10 at work..
Jack Poulson
@poulson
Nov 14 2016 00:40
I just deleted a 38 GB file in /var/log/jenkins
hopefully that isn't a problem
Ryan H. Lewis
@rhl-
Nov 14 2016 00:40
lol
Ryan H. Lewis
@rhl-
Nov 14 2016 01:00
@poulson: I just spoke with steve canon, he explained that dyld has a function stack_not_16_byte_aligned_error
which has the following assembly:
-> 0x7fffb178c506 <+0>: movdqa %xmm0, (%rsp)
0x7fffb178c50b <+5>: int3
it uses SSE to enforce alignment
he confirmed that newer gcc has these alignment issues.
he recommends reporting the elemental test case involving the cholesky factorization as a bug to gcc
and also mentioned that all OSX after 10.10 enforce this stack alignment
he also said that after 10.10 it is extremely difficult to produce unaligned memory.
Jack Poulson
@poulson
Nov 14 2016 01:02
it might be highly nontrivial to construct a Minimum Reproducible Example that doesn't depend on 10k lines of Elemental, but it might be worth doing
I'm running on 10.12 and producing this :-)
Ryan H. Lewis
@rhl-
Nov 14 2016 01:03
I pointed this all out to him, and he said that Elemental is well known enough within the software community to warrant producing a small test case that depends on it.
and that he also said it is more likely to do with the version of gcc
he said 6.2.0 is too new
Jack Poulson
@poulson
Nov 14 2016 01:04
Intel is supposedly officially using Elemental to test its compilers now
Ryan H. Lewis
@rhl-
Nov 14 2016 01:04
that’s great :)
Jack Poulson
@poulson
Nov 14 2016 01:05
and a certain well-known FEM company is supposedly transitioning to its sparse-direct solvers
I'll try building with Homebrew's gcc5 to see what happens
Ryan H. Lewis
@rhl-
Nov 14 2016 01:29
@poulson it looks like my improved BFGS is working and all the unit tests seem to be passing.
Ryan H. Lewis
@rhl-
Nov 14 2016 01:36
:)
weird. of course on my laptop it is just failing on a quadratic..
Ryan H. Lewis
@rhl-
Nov 14 2016 01:58
@poulson steve also mentioned that this bug if reported could likely be fixed easily. He suggested that it is likely some kind of fancy tail recursion optimization that works on x86 but not on 16 byte aligned systems.
Ryan H. Lewis
@rhl-
Nov 14 2016 02:32
uh, shoot, just kidding, had a bug in the test framework, it was hiding the failures that I am seeing locally. Thought it was just some stupid problem locally.
Jack Poulson
@poulson
Nov 14 2016 06:34
still an issue in GCC 5