►
From YouTube: SimPEG meeting Aug 5, 2020
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
I
will
hit
record
perfect
well
who
wants
to
start
us
off
this
week.
C
I
don't
have
much
to
add
just
other
than
yeah
so
I've
I
built
mumps
on
our
one
of
our
128
core
machines
and
did
some
testing
and
yeah
it.
It
scales
beautifully.
A
C
Yeah,
if
we
have
big
problems
using
bumps
on,
however
many
cores
you
have
and
it's
seeming
that
the
bigger
the
problem,
the
better
it
scales
at
higher
core
too
I've
got
a
little
charts,
maybe
I'll
just
I'll
post
them
on
the
dev
channel
there
on
slack
yeah
and
then
glenn
did
some
efficiency
and
yeah
we're
getting
like
80
efficiency.
After,
like.
A
C
40
some
cores
and
yeah
just
going
up
from
there.
Just
I've
been
having
problems
with
the
the
pine
bumps
rapper
I
was
digging
around.
I
don't
know,
I'm
not.
I
haven't
really
played
with
cython
too
much
so
I'm
kind
of
in
there
stumbling
around
trying
to
make
things
work
joe
went
in
yesterday
and
the
wizard
that
he
is
was
like.
Oh.
D
C
C
I'm
still
fighting
with
it.
I've
actually
got
one
colleague
on
writing
a
fortran
or,
like
the
the
f2
pie,
wrapper
for
it
yeah
just
to
see.
If
we
can
go
that
way
to
get
it
get
some
more
testing
done.
F
C
And
I
did
all
the
testing
with
frequency
domain
stuff.
So
it's
all
the
z
mumps
is
the
double
complex
version
yeah.
So
we
should
be
good
to
go
for.
B
C
You
might
not
have
to
worry
about
the
fortran
compiler
if
joe
gets
his
side
or
like
if
joe's
already
tested
and
said
it's
working,
so
I
don't
think
he
had
to
go
through
fortran
stuff
at
all.
Oh,
I
see,
but
the
other
end
the
other
angle
that
I'm
working
on
yeah
you
would
have
to,
but
the
other
cool
thing
is.
Is
we
joe-
and
I
were
joe
was
pointed
me
at
this
repositorforge
repository
and
it
looks
like
they
have
one,
that's
pretty
close
to
what
I
built.
C
It's
five
version
5.1
something
but
it,
but
I
looked
at
the
make
file
and
it
seems
to
be
using
a
lot
of
the
same
things
like
parameters
and
pc,
porch
or
scotch.
So,
yes
should
be.
It
should
be
actually
easy
to
install
too
if
you're
just
doing.
Oh.
C
A
Would
you,
when
you
think
it's
the
right
timing
to
try
this
out?
I
mean
I
would
be
super
game
to
try
this
on
google,
if
you
have
like
sort
of
a
set
of
steps
of
just
like
give
this
a
shot
at
the
high
level.
F
C
Okay,
cool
yeah.
I
have
a
recipe
for
the
one
that
I've
made
already
it's
just
joe's
having
more
success
with
the
version
5.1.
So
whatever
one
can
we
get
to
first
we'll
yeah
we'll
get
a
hold
yeah.
That's
awesome.
A
That's
really
really
exciting.
I
feel
like
that's
going
to
unlock
unlock
some
really
cool
stuff
and
actually
being
able
to
scale
and
yeah
wow
cool.
C
C
B
C
C
Some
thousand
element
model
and
partiso
took
1.46
with
access
to
those
120
cores,
and
I
got
bumps
down
to
0.22
milliseconds,
oh
wow
yeah,
but
then
again
the
partiso
isn't
really
using
all
those
cores.
So
it's
really
just
stuck
down
to
a
few.
So
it's
not
quite
the
fair
comparison.
Other
than
mums
can
use
it
all
and
if
you
can
use
it
all,
it's
fast.
F
C
In
that
case
yeah,
oh,
so
in
that
case
I
I
got
what?
Where
was
that
number
point?
Oh
no
wait!
Oh
no,
sorry,
I
forgot
a
zero
there.
Okay,
no!
I
was
reading
my
wrong
chart
so
at
128
it
solves
in
point
zero,
five
milliseconds,
whereas
with
one
core,
it's
0.22,
so
it's
even
faster
on
like
the
single
core
than
party.
So.
B
Oh
wow,
that's
great
and
were
you
using
like
a
2d
mesh
or
a
3d
mesh.
C
3D
mesh
yeah
there's
only
about
56,
000
elements
or
sorry
yeah,
I
think
elements
or
shells.
Sorry,
that's
great.
B
That's
great:
have
you
tested,
like
kind
of
maximum
size
that
you
could
run
like.
C
That's
what
I
was
hoping
to
like
with
the
wrappers.
We
can
have
better
access
to
that,
because,
right
now,
if
I
do
something
with
pardeso,
I
have
to
write
it
to
like
a
ascii
file
so
that
I
can
bring
it
into
the
mumpside
and
as
soon
as
you
get
into
like
a
hundred
thousand
cells,
it
takes
like
a
day
to
write
that
ascii
file.
So
so
I
was
hoping
to
do
the
wrappers
and
then
do
the
full
test
on
the
large
problems
say
the
cscm
one
there.
A
Yeah
I
mean
the
base
mesh,
I
think,
is
like
9
million
cells,
so
I
was,
of
course,
going
to
reduce
that
with
the
octree,
but
it's
still
probably
going
to
be
on
the
order
of
several
million
cells,
so
that
would
be
an
ultimate
test
for
sure.
B
A
But
that's
also
like
I
mean
a
completely
uniform
mesh,
so
there's
stuff
we
can
do
to
trim
that
down,
but
it's
still
gonna
end
up
being
several
probably
at
least
two
million
cells.
So
right
we'll
see
what
we
can
do
with
that.
A
Yeah
yeah,
so
I've
been
running
it
on
google
and
you
can
get
a
high
memory.
Node!
That's
128,
gigs
of
ram,
so.
B
C
D
Oh
yeah
congrats,
well
yeah.
It
got
accepted
for
this
for
this
time
for
the
like
the
jgi
paper,
like
for
the
joint
methodology,
so
I
will
add,
send
all
the
things
this
afternoon
doesn't
seem.
We
got
any
revision,
I
don't
know
even
if
it
seems
like
the
adhesive,
just
looked
at
it
and
said
like
yeah,
that's
fine.
D
Yeah
so
yeah,
I
haven't
spent
much
time
on
simpeg
the
last
week.
I
was
mostly
like
yeah
reading
papers
and
like
following
the
some
of
the
tutorials
with
jempai.
I
was
actually
also
looking
at
the
as
a
material
that
editor
put
online
of
through
3m
by
moto,
and
that's
pretty
good
too,
as
like.
He
has
two
like
some
serious
geological
models
there
that
so
yeah.
D
I
was
looking
at
that
and
I'm
thinking
too,
like
I
took
more
question
at
the
moment,
but
like
the
so
I
will
probably
be
defending
early
september,
and
it
seems
like
a
real
house
like
to
get
like
an
odms
to
the
it's
gonna
be
online
for
the
for
the
defense
and
it
seems
a
real
ass,
always
a
great
studies
to
get
the
an
audience
etc.
D
There
like
they
share
it
only
with
the
committee
and
the
and
the
defendant,
and
you
need
to
request
individually
for
to
be
send
out
the
link
etc
and
and
they
and
they
they
handle
the
the
the
grass
that
is
handling
the
zoom.
Apparently
so
I
don't
think
they
are
recording
either
so
so
I
was
thinking
that
anyway.
I
would
probably
like
to
do
like
a
like
a
fake
run
before
doing
the
defense
like
run
through
the
slide,
and
to
have
like
some
feedbacks
before
the
before
that.
D
So
I
was
thinking
either
we
do
that
as
a
simple
as
a
simple
meeting
or
at
some
other
point,
and
if
anybody
is,
is
willing
to
to
be
an
audience
for
for
like
a
a
personal
training
for
the
for
the
defense,
and
then
we
can
record
that,
as
that
thing
as
like
to
keep
a
track
of
that
of
that
presentation,.
A
Yeah
for
sure
we
can
definitely
either
take
a
simpeg
meeting
or
do
something
on
a
different
day,
whatever
whatever
works
so
feel
free
to
just
like
claim
a
meeting
as
yours
when,
when
you're,
when
you're
ready
to
do
that.
D
B
Sounds
pretty
cool?
I
would
be
very
interested
to
see.
D
Yeah,
I
can
share
the
tests
and,
if
you,
if
you
like
the
pdf
of
the
days,
if
you
wish
it's
it's
on,
google
drive
like
it's.
Hopefully
I
will
have
not
too
many
connections
on
the
testis.
Anyway,
all
all
all
chapters
will
have
been
published
by
the
time.
I
get
my
correction,
so
that's.
F
Well,
I
think,
for
me
I
was
working
work,
work
and
taking
some
vacation,
that's
kind
of
that
time
of
the
year
right.
F
I
just
went
back
from
banfield
with
chantal
was.
B
B
F
B
Now
I
got
a
little
bit
more
time
and
yes,
I'm
looking
forward
to
revisit
the
devin's
reflector
for
1d
code
and
also
yeah.
I
want
to
spend
more
time
these
days
to
kind
of
package
up
what
we
developed
here
in
stanford,
so
yeah,
I'm
looking
forward
to
have
a
bit
more
time
in
these
days.
B
So
yeah,
that's
that's
where
I
am,
and
one
thing
that
I
was
thinking
was
developing
a
2d
empty
code
because,
like
feels
like
I
mean
we're
pushing
through
to
develop
the
3d
code,
but
that,
in
terms
of
thinking
about
the
use,
2d
code
could
still
be
very
useful.
B
C
B
Right
yeah,
so
it
seems
like
an
interesting
avenue
and
I
don't
know
on
my
end,
I
mean
developing
to
the
empty
code
itself
is
not
very
interesting
in
that
research
side.
I
mean
it's
done,
but
I
guess,
if
you
think
about
combining
with
other
things
like,
like
tivo's,
joint
inversion
for
instance,
and
having
a
tool
that
set
like
that,
will
be
very
helpful
and
it's
actually.
It
is
fast.
C
If
you
need
a
hand,
soggy
I'll
definitely
put
some
time
to
help
for
sure.
E
Well,
always
having
like
a
lower
dimensionality
inverse
code
is
good
right.
I
mean
look
how
valuable
the
1d
code
is,
for
just
you
know,
analyzing
data
quality
control,
getting
a
first
pass
about
things
and
same
is
true
with
the
mt
and
there's
still
a
lot
of
data.
That's
just
acquired,
you
know
with
lines
of
of
mt
data,
and
the
other
thing
is:
if
it's
you
know,
then
you
can
also
experiment
with
you
know
boundary
conditions
right,
so
I
think
it's,
I
think
it's
a
nice.
B
E
A
B
I'm
like
when
I'm
thinking
about
myself
as
a
an
invert
like
a
running
an
invert
person
who
are
running
an
inversion
having
a
fast
code
is
very
important
because
I
mean
if
you
need
to
wait
a
couple
days
to
run
inversion
and
your
mindset
is
gone.
So,
if
you
have
like,
let's
say
like
a
couple
minutes
to
run
the
inversion
and
get
a
result
and
think
about
what
it's
supposed
to
be,
and
what's
your
sort
of
expectation
and
change,
the
parameter
that
process
like.
I
think
that
speed
is
very
important
in
that
process.
B
Otherwise
you
lose.
You
lose
kind
of
momentum.
I
thought
so.
Yes
yeah,
I
think,
still
lower
dimension
code
kind
of
has
a
value
in
in
in
that
way.
Anyway,.
E
Not
really,
but
we're
pleased
that
we've
got
all
of
the
stuff
in
with
the
gwb.
So
thanks
everybody
who
contributed
to
that
so
that's
kind
of
nice
to
get
to
get
that
off
and
now
we
can
concentrate
on
we've
got
some
presentations
at
seg
and
agu
actually
presentations.
That
scg
must
was
that
middle
of
august
that
we
were
supposed
to.
E
Okay,
that's
that's
good,
yes
and
that's.
Well,
that's
sort
of
been
the
main
occupation
after
this
is
after
this
is
over
soggy
and
dom.
Maybe
we
can
continue
chatting
about
the
lp
lq
inversion
to
kind
of
look
at
that
other
paper,
but
it's
kind
of
related
to
synthetic
but
different
matters.
So
we
might
report
back
later.
F
On
on
the
sag
front,
so
yeah-
I
don't
know
if
you
guys
saw
the
emails
flying
by,
but
sorry
cg
finally
got
it
to
reason
and
decided
to
go
full
full
online
right.
So
jenny
cole
sent
an
email
to
all
the
post
convention
people
and
said
that
they
will
review
how
they
want
to
do
this,
and
if
it
looks
like
it's
going
to
be
a
full
zoom,
probably
interactive
one.
So
in
the
end
after
fighting
with
them,
we
might
you
might
get
what
we
wanted
in
the
end.
F
So
we'll
see
we'll
just
wait.
The
they
promised
an
update
with
them
within
a
few
days.
F
Sorry,
I'll
I'll,
I
can
I'll
forward
the
email,
the
emails
that
I
got
and
then
then
it
came
out
last
week.
I
believe
last
week,
yeah.
E
A
Yes,
on
my
end,
I've
been
working
on
sort
of
refining
the
layered
model
for
the
paper
with
theater,
and
so
I
actually
I'm
gonna,
send
that
hopefully
this
afternoon
and
if
folks
are
willing
to
take
a
look
if
anyone's
willing
to
take
a
look
and
see
if
there's
improvements
you
think
could
be
made,
I
would
certainly
appreciate
input.
A
I
basically
got
it
down
to
the
point
where
we're
sort
of
like
median
three
percent
error,
which
is
pretty
good,
but
I
think
we
should
still
be
able
to
drive
it
down
further,
and
so,
if
anyone
has
ideas,
when
you
look
at
the
mesh
or
the
interpolation
or
things
like
that,
I'd
be
happy
to
sort
of
walk
through
what
I've
done,
but
would
would
appreciate
a
fresh
set
of
eyes
on
it,
because
I've
been
staring
at
it
twiddling
twiddling
knobs
to
see
what
I
can
do,
but
yeah
another
another
perspective
would
be,
would
be
appreciated.
B
What
like
a
it's,
a
median
three
percent
seems
like
like,
if
you
have
like
outliers,
like
pretty
large,
like
like
just
just
curious,
what
like
the
median
seems
like
we
usually
compute
the
mean
or
l
square,
like
rms
error,.
B
A
And
so,
like
the
mean
you're
gonna
have
some
some
of
those
are
gonna,
be
just
artificially
shifted
from
that.
So
there
aren't
huge
outliers,
but
it's
really
those
zero
crossings
that
cause
the
outliers.
So.
A
Yeah,
so
yeah
so
I'll
send
that
around
hopefully
this
afternoon
and
then
we'll
be
sort
of
working
on
getting
the
next
example,
which
will
be
this
marlin,
one
which
would
be
pretty
big,
I'm
going
to
start
by
just
like
trying
to
get
a
really
coarse
volume,
average
version
of
it
just
up
and
running
so
that
it
got
kind
of
the
workflow
in
place
and
then,
hopefully,
by
the
time
I
pop
that
button,
then
maybe
the
mumps
will
all
converge
and
be
able
to
be
able
to
start
running
some
big
examples
there,
and
then
I
guess
some
of
the
other
things
so
on
the
swung
slack,
there's
again
been
sort
of
the
suggestion
of
thinking
through.
A
A
My
sort
of
initial
suggestion
would
be,
I
think,
in
some
ways
it
might
be
kind
of
cool
to
merge.
Some
of
the
methods
focused
channels
so
like
have
em,
be
back
and
forth
and
have
single
field
be
back
and
forth
and
sort
of
more
more
places
where,
like
that
yeah,
it's
a
methods-based
conversation
at
this
point.
A
I
don't
know
if
we
want,
like
the
meetings
stuff
or
the
dev
channel
on
there,
but
there
might
be
some
things
to
think
through
just
like
how
how
we
might
want
to
merge
or
like
open
up
some
more
communication
lines
with
that
community.
A
So
there's
a
couple
things
we
could
do
one
we
could
just
get
rid
of
the
em
channel
on
simpeg
and
then
just
have
those
conversations
and
so
on.
The
other
thing
that
we
can
do
is,
I
think,
there's
a
concept
of
joint
channels
between
slack
groups,
and
so
it's
something
is
sent
in
the
simpeg
em
slack.
It
would
show
up
also
in
the
swing
slack
because
they're
like
one
in
the
same
thing,
which
might
be
nice
yeah,
yeah,
yeah,.
A
Okay,
yeah.
A
To
think
about,
and
we
can
maybe
revisit
other
things
well,
I
guess
joe,
do
you
have
any
hey.
G
G
That
they
have
looking
into
it
more
and
looking
at
more
how
it's
interfaced
and
enter
how
it
interacts
with
things
it.
Unfortunately,
we
will
not
be
able
to
like
use
it
we're
going
to
have
the
same
problems
with
desk
interacting
with
it
than
we
did
and
that
we
had
with
our
diesel.
G
G
Of
built
in
that
similar
manner
and
that
okay,
you
have
a
whole,
you
have
one
long
object,
one
like
one
structure
that
has
a
bunch
of
pointers
to
things
to
it
and
that's
what
you
pass
around
everywhere,
so
what's
actually
happening
with
mumps
is
that
okay,
there's
one
object.
That
has
a
pointers
to
every
variable
that
you
need
to
pass
to
the
functions
that
operate
on
it.
G
So
when
you
have
to,
when
you
solve
a
right
hand,
side,
you
have
this
struct
this
variable,
or
this
like
object.
That
represents
everything
that
you
pass
to
it.
So
when
you
factorize
it,
you
give
it.
The
I
j
is
a
values
of
the
matrix
right
and
then,
when
you
solve
it,
that
same
object
also
has
like
right
hand,
side,
pointers.
G
G
B
G
G
Kind
of
thing
is
that
multiple
calls
to
the
solver
to
the
solve
operation
from
this,
from
different
threads
of
the
same
process
for
causing
clashes
like
memory
clashes
and
overwriting
stuff
and
thinking
back
to
more
about
how
that's
implemented
and
it's
basically
implemented
the
exact
same
way
like
you
assign
pointers
to
things
I
don't
know,
it
just
seems
like
a
it's
like
a
leftover
of
whole
90s
implementations
of
things
before
they
were
well
like
for
object.
Orientedness
got
around
everywhere,
I'm
not
sure.