►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Deep Customized Kubernetes for Machine Learning in Tencent - Shengbo Song, Tencent
The Kubernetes and Tensorflow are playing an important role in machine learning, However, the original Kubernetes doesn't serve machine learning well, lacks of accurate GPU scheduling policy, GPU topology awareness and resource limitation etc. In this presentation, we will review recent events about machine learning in Kubernetes community, give us a overview of what changes have been made to Kubernetes to support machine learning in Tencent, the challenges it faces, and look forward to how to address the challenges.
To learn more click here: https://sched.co/FuJi
A
B
A
C
A
B
Community
version
1.6
to
1.1,
which
is
about
to
be
released.
We
keep
focusing
on
the
support
of
the
GPU.
We
realize
that
community
has
always
been
dedicated
on
the
extension
of
API,
but
they
didn't
really
focus
too
much
on
special
special
drops
and
doesn't
have
certain
jobs
as
amazing,
for
example,
a
GPU
topology
awareness
and
host
device
for
some
apps,
because
some
trying
to
use
some
devices
on
the
host,
but
they
don't
have
the
necessary
support
and
also
resource
management.
C
C
A
A
C
B
C
B
C
A
C
B
B
You
transmit
a
data,
one
gigabytes
for
bigger
price
and
ATM
right.
When
you
pick
GPU
zero
and
GPU
one.
You
can
see
the
time
for
GPU
zero
and
one
is
only
half
of
the
time
used
by
GPU
0
&
3,
so
it
can
selectively
perceive
this
GPU
topologies
or
make
you
faster
to
transmit
data.
When
the
user
is
using
Gaia
stack
platform,
they
don't
have
to
choose.
The
GPU
modular
can
automatically
pick
the
best
GPU
topology.
At
that
moment,.
A
A
B
C
B
C
B
Have
realized
certain
GPU
sharing
approaches,
for
example,
for
TVM
there's
a
la
vida,
great
and
there's
a
multiple
process
use
mps
surveys.
We
have
looked
into
this
two
approaches
by
the
media.
We
have
done
some
tests
as
well.
We
approach
by
Nvidia
has
a
good
official
support
and
good
satellite
appointment,
so
the
users
can
easily
use
the
GPU
sharing,
but
in
some
cases
or
at
some
extent
it
is
not
suitable
for
kubernetes
platform
or
kaya
stack
platform
because
of
the
following
for
a
media
grid.
B
Ease
and
withdraw
it
can
affect
the
MPS
server
and
clients
as
well.
So
one
program
crashes
will
also
lead
to
the
downtime
of
other
applications
as
well,
and
we
don't
want
that
and
resource
limit
is
another
reason.
The
Nevada
has
the
latest
motor
series
GPU.
Officially,
they
said
you
can
limit
a
MPs
client
and
they
can
use
multiple
computing
threads,
but
also
in
the
documents
they
said.
Will
one.
A
B
One
application
started:
it
cannot
be
modified
so
when
there
are
two
users
when
they
are
using
0.3
GP
resources
together,
respectively
and
then
only
have
used
0.6
this
resource
and
if
one
user
doesn't
use
the
computing
imperative
capacity
obedience,
hard
limits
will
lead
to
such
a
result
where
the
user
could
have
fully
used
the
computing
resources.
However,
in
reality
they
can
only
use
their
0.3.
So
what
we
want
is
well.
There
is
nobody.
C
A
B
To
each
MDS
client,
but
when
they
submit
a
BP
to
the
setting,
they
can
just
be
like
a
CPU.
They
can
set
their
point
3/4
or
0.75.
But
if
you
use
the
same
approach
to
sense
Nvidia,
they
will
not
follow
what
you
or
you
expected.
They
will
not
give
user
a
0.3
or
user,
be
there
1
7.
Actually
it
will
distribute
that
evenly.
D
D
C
D
D
C
D
C
D
C
C
D
C
C
C
A
D
C
D
A
C
D
D
D
C
C
D
D
C
D
B
B
So
what
if
Seth
has
any
advantages
or
anything
special
about
it?
So
that's
the
question
and
our
self
is
maintained
by
a
dedicated
team
and
in
Tencent
we
have
OCR
project.
We
would
actually
do
some
optimization,
which
will
support
a
high
IOPS
file
storage.
So
it
will
be
a
big
improvement
comparing
to
the
default
of
the
community.
So
when
there's
when
they
read
a
large
number
of
small
files,
performance
will
not
drop
dramatically
and
the
tenant
cloud
will
also
use
our
technology.
A
B
A
A
B
A
B
A
A
B
Performance,
we
have
compared
the
performance
as
well.
Actually
those
cards
is
distributed
to
this
VM
and
then
he
cannot
be
distributed
to
the
others.
So
you
know
scheme
we
realized
and
sharing,
and
it
is
a
lightweight
server.
So
not
all
GPU
commands
will
go
through
our
server.
We
only
selects
certain
GPU
commands.
B
We
actually
issue
release
their
paper.
The
related
technologies,
which
will
be
released
in
October,
which
was
released
in
October,
was
something
it
was
quite
detailed.
Introduction
I
can
send
you
the
link
of
that
paper,
so
you
can
take
a
look
and
see
how
we
realized
our
thread
limits
and
times
life.
A
C
C
B
A
B
A
A
C
B
B
But
in
this
case,
if
you
realize
the
model
is
not
good,
as
expected,
the
user
might
realize
there
are
some
changes
and
then
they
will
feel
why
this
sudden
suddenly
there's
a
change,
but
developers
don't
want
users
to
feel
such
a
change.
So
in
this
case
you
can
use
the
tab.
You
can
use
the
image,
change
the
image
of
a
few
parts
and
update
it
to
the
testing
version,
and
then
it
can
connect
the
testing
version
and
the
ordinary
version.
B
B
Let
me
repeat
his
question
so:
for
their
community:
it's
plugging
the
user
plugin
to
support
the
invidious
GPU
by
the
same
set
in
media.
Don't
have
enough
inputs
in
it,
so
the
only
distributed
on
card
and
for
us
we
provide
the
GPU
sharing.
So
his
question
is:
is
it
possible
that
we
would
actually
open
this
management
model?
I
might
not
be
able
to
answer
to
you
correctly,
but
based
on
our
perception,
open
perception.
If
we
feel
it's
appropriate,
we
will
make
it
open
source.
B
We
also
issued
an
essay
on
this
technology,
so,
if
you're
interested,
we
can
send
you
the
paper
for
the
conference,
so
I'll
tell
you
when
the
paper
is
going
to
be
released
because
the
paper
is
already
selected
by
the
meeting,
so
it
is
only
a
time,
a
matter
of
when
it's
released.
Thank
you
very
much
for
your
listening
to
today's
speech
and
that's
the
end
of
it.
Thank
you.