►
From YouTube: 2021-09-14 - Challenges and successes with a hybrid multicloud implementation for research computing
Description
NERSC Data Seminars Series: https://github.com/NERSC/data-seminars
Speaker:
Jonathan Skone
Advanced Technologies Group
National Energy Research Scientific Computing Center (NERSC)
Lawrence Berkeley National Laboratory
Title:
Challenges and successes with a hybrid multicloud implementation for research computing
Abstract:
A review of the efforts to evolve an academic on-premises HPC ecosystem to the cloud through the hybrid multicloud solution named Skyway, will be presented. Skyway incorporates multicloud computing resources as elastic extensions of its on-premises HPC infrastructure and makes use of a system-level software package to interface the job scheduler and cloud SDKs, resulting in a seamless experience for users when interacting with both on-premises and cloud resources. The implementation is general enough to interface with one or more cloud providers, which currently include both Amazon AWS and Google GCP. The challenges encountered and the use cases where it has been successful, will be elaborated.
Bio:
https://www.nersc.gov/about/nersc-staff/advanced-technologies-group/jonathan-skone/
Host of Seminar:
Nicholas Wright
Group Leader, Advanced Technologies Group
National Energy Research Scientific Computing Center (NERSC)
Lawrence Berkeley National Laboratory
A
A
A
B
So
the
the
title
of
the
talk
is,
you
know,
challenges
and
successes
with
a
hybrid
multi-cloud
implementation,
and
so
this
is
something
that
a
lot
of
research
computing
centers
have
been
pushing
forward
with
thinking
about
and
trying
to
implement,
either
by
pulling
in
vendors
and
using
ready-made
offerings
or
trying
to
implement
something
on
their
own.
So
this
this
activity
started
several
years
ago
and
we
had
just
recently
submitted
a
short
paper,
a
conference
paper
to
the
the
perk
conference
that
was
held
in
july.
That
references
is
here.
B
So
I,
before
kind
of
getting
down
to
the
use
of
the
cloud
resources
at
the
university
of
chicago.
I
should
kind
of
preface
that,
by
giving
you
a
picture
of
the
landscape
of
of
research
computing
at
the
at
a
at
the
university
of
chicago,
so
the
traditional
hpc
is
still
a
dominant
driving
force
for
scientific
applications
with
the
the
environment
consisting
of
more
than
1800
compute
resources,
they're
heterogeneous
resources.
So
there's
there
are
several
clusters
that
are
supported
and
those
clusters
themselves
are
heterogeneous.
B
B
Instead,
they
are
dedicated
for
your
use
and
only
your
group's
use,
and
so,
if
you
don't
use
it,
it
just
sits
idle,
the
but
the
most.
For
the
most
part,
the
people
that
do
pay
for
and
have
dedicated
resources
have
them
in
use
more
than
85
percent
of
the
time.
The
the
resources
that
are
consumed
are
mostly
by
the
folks
that
are
doing
things
that
I
I
classify
as
non-traditional
hpc
in
the
sense
of.
B
If
you
went
to
a
do
resource
at
dewey
or
other
national
facility,
you
have,
or
at
least
the
the
bigger
oscar
facilities
like
argon
and
oak
ridge.
You
have
a
very
high
barrier
to
using
those
resources
and
that
you
have
to
demonstrate
scalability
and
you
have
to
be
able
to
run
on.
You
know
hundreds
if
not
thousands
of
nodes,
and
so
for
most.
B
You
know
that
that's
kind
of
like
the
the
entry
point
for
doing
hpc
on
those
resources
which
we
don't
typically
have
at
an
academic
research
computing
center.
There
are
very
highly
scalable
jobs,
it's
just
that
they
don't
that
they're
less
than
20
of
the
the
total
utilization,
so
that
that
doesn't
mean
that
they
don't
consume
a
fair
for
a
number
of
service
hours.
B
How
do
I
do
that
so
ways
that
the
the
university
computing
center
has
adapted
to
support
these
evolving
needs?
There
are
a
number
of
them
so
one
in
in
terms
of
providing
web
portals
and
data
gateways.
There
is
rancher
service.
It's
it's
a
little
bit
dated
this
is
this
is
rancher
one
with
which
which
has
the.
B
Is
is
with
with
cattle,
not
kubernetes,
that
is
integrated
with
git
lab,
so
this
self-hosted
git
lab
that
that
has
that
talks
to
rancher
through
a
registry
where
the
privileged
staff
and
also
privileged
developers
can
push,
can
can
push
images
to
this
registry
that
the
ranchers
then
able
to
provision
those
services.
So
this
provides
access
to
web
services
and
scientific
gateways.
B
But
it's
unfortunately,
unlike
at
nurse,
it's
it's
something
that
is
only
accessible
to
the
staff.
So
the
staff
have
to
provide
this
service,
and
so
it's
it's
only.
It's
only
provisioned
on
a
case-by-case
basis,
depending
on
the
particular
projects
and
and
the
visibility
of
it
will.
Will
there
be
given
a
the
staff's
time
to
stand
up
those
services.
B
So
there's
been
the
support
for
containers
in
hpc
and
and
that's
that's
been
in
existence
through
sometime,
like
2017
and
that's
that's
been
that's
been
provided
through
singularity
and
there's.
There's
document,
there's
minimal
amount
of
documentation
for
folks
to
get
up
and
running
using
singularity
on
these
resources,
and
there
is
a
separate,
ci
cd
server,
which
is
meant
to
provide
privileged
staff
and
developers
with
an
environment
to
create
docker
containers.
So
this
is
not
for
the
average.
B
This
is
not
for
the
the
regular
sets
of
users
of
the
hpc
resource.
It's
it's
for
the
folks
with
admin
privileges
to
create
containers
in
that
environment,
as
well
as
the
developers
of
a
particular
project
to
have
access
to
and
create
containers
that
can
be
used
on
the
hpc
resource.
So,
for
the
most
part,
users
themselves
have
to
bring
their
own
containers
to
to
run
on
the
hpc
resource.
B
It's
a
it's
a
cloud
agnostic
gateway
infrastructure
and
software
package
that
provides
services
that
essentially
links
the
on
premise
and
cloud
resources,
and
another
thing
that
is
more
recently.
This
is
just
sort
of,
as
I
was
exiting
the
university
of
chicago
more
folks
that
are,
in
the
I
say,
not
non
non-traditional.
B
B
So
globus
is
used
in
all
cases
to
mediate
data
transfer.
There
is
a
ftp
server,
also
sorry
http
server,
also
running
for
for
the
qrs
project,
just
to
make
it
accessible
to
quickly
be
able
to
grab
files
without
having
to
set
up
any
kind
of
transfer
transfer
credentials
to
to
move
data
by
via
globus.
So
an
example,
the
example
here
of
the
xrom
data
management
portal-
this
one
is-
is
used
there's
this
is
using
the
the
the
globus
sdk
to
create
a
data
portal
for
ingestion
of
data
with
a
specific
schema.
B
B
B
So
the
support
for
containers
in
hpc
environments,
we
use
singularity
it's
the
most
prevalent
for
research
computing
and
it's
you
know
it
has
different
approach.
Security
model
compared
to
docker.
The
use
of
the
singularity
at
the
computing
center
essentially
enables
folks
to
use
docker
containers,
so
they
they
can
build
with
docker,
but
they
they
convert
and
run
their
images
as
as
singularity
image
files.
B
But
the
thing,
though,
is
that
users
cannot
build
in
in
this
environment,
and
that
is
it's
just
a
small
pain
point
for
some.
For
the
most
part,
a
majority
of
the
users
just
simply
need
to
be
able
to
pull
an
existing
docker
image
that
has
been
created
by
a
particular
software
project.
So
they're
not
they're,
not
interested
in
building
their
own
containers.
B
Because
that
can
you
know
in
my
experience
that
can
be
time
consuming
or
you
know
it
can,
instead
of
spending
just
a
few
hours
or
getting
something
done
in
a
day.
You
end
up
spending
a
week
until
you
get
the
right
people
involved
and
that
you
know
this.
This
also
can
can
lead
to
a
delayed
resolution
of
any
kind
of
software,
but
so
the
the
intent.
B
So
in
particular,
we
were
interested
in
using
this
at
alcf,
and
this
is
the
this
is
the
particular
software
that
was
was
containerized
and
and
for
the
scaling,
scaling,
application
and
just
sort
of
in
a
nutshell.
What
was
done
in
order
to
use
this
at
two
different
resources,
so
the
the
I
mentioned
before:
there's
an
environment
for
developers
and
for
privileged
users
or
not
users,
privileged
staff
at
the
university
of
chicago
to
build
docker
containers.
B
B
B
When
you
go
from,
you
know
the
small
research
computing
center,
where
you
have
hundreds
of
nodes
available
to
thousands
of
nodes,
there's
no
there's
no
drop-off
and
performance
or
penalty
and
performance
for
running
in
the
container
on
on
on
the
lcs
facilities.
So
the
this
plot
here
shows
both
intel
and
gcc
builds
and
effectively.
The
intel
build
is
is
more.
A
B
So
the
the
intent
of
using
these
containers
was
again
for
the
assessment
of
portability,
because
it's
something
that
you
know
users
at
the
university
of
chicago,
which
also
have
access
to
resources
at
doe
facilities
such
as
argonne
and
nurse
they
want
to.
They
want
to
assess
you
know
whether
they
can
have
a
small,
smaller
environment,
to
develop
in
and
be
able
to
run
and
and
be
able
to
push
out
to
larger
resources
where
well,
where
they
have
particular
workloads
that
or
particular
job
sizes
that
that's
appropriate.
B
So
the
you
know
that
this
can
also
be.
You
know,
thought
in
terms
of
the
the
you
can
think
about
this
in
terms
of
fare
and
the
reproducibility
aspect
of
using
containers
for
for
reporting
across
resources,
because
all
these
packages
that
are
contained
within
the
container
are
accessible
or
reusable
across
the
various
resource
centers.
B
B
The
the
portability,
though,
requires
that
the
glibc
and
the
kernel
have
compatibility
between
the
hosts
and
the
container,
and
that
this
I
think
a
lot
of
this
has
been
since
is
not
really
as
much
of
an
issue
now.
As
it
was
say,
this
is
a
year
and
a
half
ago
that
the
the
there
were
there
are
a
lot
of
it
there.
There
were
some
hurdles
with
using
the
mpi
gpu
drivers
and
having
api
compatibility
between
them.
A
B
Set
of
of
containers
and
and
and
and
doing
this
reproducing
the
same
study
across
various
sites,
so
the
the
I
guess,
the
biggest
takeaway
from
doing
this,
because
I
I
don't
know
that
this
is
this
is
novel.
The
folks
at
lionel
had
also
done
a
scaling
test
around
the
same
time
that
that
we
had
done
this,
but
they
they
had
they
had
published.
B
B
B
They
did
not
want
to
pay
for
dedicated
resources,
because
that
would
be
a
waste
of
of
their
of
their
money
because
they
only
need
maybe
two
months
out
of
a
year
where
they
they
need
to
fully
use
several
hundred
nodes
in
order
to
run
a
workload
and
so
for
them.
Having
you
know,
a
having
high
resource
usage
but
infrequently
was
was
a
a
a
need
for
them
to
be
able
to
pay
on
on
demand.
B
For
for
these
resources,
there
are
also
faculty
that
don't
want
to
manage
user
accounts
or
any
of
the
resources.
Provisioning
resources.
That's
just
a
headache
in
general
for
folks
to
manage
any
of
the
the
cloud
resources
themselves.
So
they
have
credit.
So
a
lot
of
the
faculty
come
into
the
university
with
their
own
startup
credits
from
aws
or
gcp,
and
some
of
them
have
billing
accounts
already,
or
they
want
to
have
a
billing
account.
B
But
again
they
don't
want
to
manage
the
users
or
have
to
deal
with
aside
from
just
understanding
where
what
what's
been
spent
and
sort
of
projecting
what
what
will
be
spent.
They
don't
want
to
handle
any
of
the
other
logistics
of
that
so
users.
So
there's
also
this
need
at
the
university
to
have
new
technologies
that
aren't
available,
on-prem
and
again
coming
back
to
community
of
users
that
don't
have
the
budget
to
buy
dedicated
resources.
B
Users
also
wanted
privilege
privileges
or
have
the
ability
to
have
privilege
to
install
software
in
an
environment,
and
that's
that's
one
case
for
which
they
come
to
us
and
ask
for
use
of
cloud
resources
and
the
other
is
they
want
to
manage
their
own
persisting
services.
So
they
want
to
stand
up
and
their
own
web
application
or
they
want
to
stand
up
their
own
database
and
they
want
to
be
the
the
sort
of
the
manager
of
that
resource.
B
And
so
the
latter
two
are
things
that
we
decided.
We
kind
of
support
we
offer
support
through
the
through
rancher
and
in
particular,
use
cases
that,
if
they,
if
there's
a
project
that
is
deemed
fit,
we
we
can,
we
can
provide.
You
know
persisting
services
or
or
web
portals
or
web
applications
for
the
user,
the,
but
the
and
the
use
of
the
privileged
environment
is
something
that
we
just
simply
don't.
B
Support
directly
folks
can
build
their
own
containers
and
bring
those
if
that's
a
requirement
to
build
a
particular
software
stack.
So
it's
not
that
you
know
we
we
aren't
making
there
isn't
an
ability
to
do
that,
so
users
do
have
the
have
the
ability
to
build
their
own
containers
where
they
can
have
control
of
the
the
environment
itself.
A
B
On-Prem
hpc
cluster
and
cloud
services,
so
it
simplifies
the
usage
to
the
cloud
computing
for
the
existing
computing
center
users.
So
that
was,
you
know,
sort
of
requirement.
One
was
I
don't
want
to
manage
like
as
a
pi.
I
don't
want
to
manage
users.
Can
you
please
set
up
something
that
is
automatically
does
this.
B
For
me,
the
the
the
skyway
also
provides
a
custom
set
of
services
and
tools
as
a
python
package,
and
these
these
tools
and
services
that
act
to
bridge
between
the
on-prem
and
the
clouds
cloud
cloud
environment.
They
they
facilitate
the
management
and
the
and
the
monitoring
payment
of
the
cloud
resources,
so
you
can
think
of
skyway
as
as
two
things.
B
B
B
It
was
not
intended
or
architected
with
the
the
idea
of
building
something
that
is
point
and
click
for
them
to
use
cloud
resources,
so
it
is
not
in
the
it
is
not
something
like
ronin
or
or
the
the
folks
that
that
that
provide
the
hpc
finite
element
codes
with
this
point-and-click
container
technology
to
use,
use
cloud
resources,
the
sorry
I'm
thinking
of
rescale
yeah.
So
it's
not
it's
not
like
rescale,
and
it's
not
like
ronins
there's,
not
this.
B
B
B
So
it's
meant
to
also
make
it
easy
for
the
users
to
to
be
added
through
the
pi's
cloud
account
with
no
management
of
users
accounts
required
from
the
pis,
there's
no
privilege
needed
to
be
imparted
from
the
pi
themselves
and
the
features
like
the
billing
there's
a
billing
module
that
prevents
spending
of
a
particular
pi
or
project
beyond
our
preset
budget.
So
I
have
a
picture
here.
You
shin
pang,
who
was
the
principal
architect
behind
the
initial
push
for
for
skyway.
B
B
B
So
the
the
cloud
scratch
and
on-premise
are
made
accessible
on
this
on
these
research
on
these
skyway
resources,
so
that
the
users
themselves
can
migrate
data
between
the
on-prem
and
the
cloud
resource
and
the
users
they
can
access
this
node
either
publicly.
They
can
go
to
the
the
login
notes
of
the
hpc
cluster
and
then
connect
to
the
skyway,
the
skyway
resource
login,
or
they
can,
if
they're
on
the
campus
network,
this
this
login
notice
is
accessible.
B
B
So
again
it
has
the
same
operating
system
and
and
environment
that
you
would
find
here
on
the
login
node,
with
the
exception
that
the
the
user's
home
is
not
the
same
as
on
the
main
hpc
space,
and
that's
because
the
the
home
here
on
this,
this
login
node,
is
not
exported
to
the
each
of
the
the
cloud
instances.
So
the
users
have
to
actually
work
out
of
a
different
space
which
is
mounted
at
cloud,
and
this
this
this
space
is
is
accessible.
B
Is
it's
the
it's
the
scratch
space
of
the
cloud
provider?
That's
mounted
on
the
login
node?
It's
they
have
to
work
out
of
that
space.
So
this
kind
of
this
can
be
a
hang
up
for
some
folks,
because
people
that
are
using
condo
environments
have
a
habit
of
you
know
managing
those
in
their
home
directory.
So
they
have
to
you
know,
point
to
a
different
path
in
order
to
install
and
manage
their
environments.
It's
it's
again
it
for
for
folks
that
are
regularly
regularly
using
and
can
manage
their
kind
of
environments.
B
In
the
cloud
that
the
folks
commonly
use
like
python
are
things
like
bromax
and
other
md
software
are
are
all
accessible
so
that
the
the,
when
I
say,
mirrored
environment
here
for
the
login
node,
the
linux,
the
linux
users
and
groups,
are
again
pulled
from
the
ldap.
So
it
is
using
the
same
credentials
to
authenticate
users
and
the
the
slurm
is
the
same.
Sets
of
commands
they're
familiar
with
it's
the
same
type
of
job
scheduler.
But
this
is
a
separate
instance
of
slurm.
B
The
modules
and
software
are
not
the
same,
but
again
we're
using
lmod
just
the
same
as
in
the
login.
So
it
has
the
same
feel
and
they
can
find
some
of
the
common
modules
that
they
were
using
on
the
hpc
resource
there
and
the
project
storage,
where
that
just
kind
of
like
the
community
file
system
at
nurse
is
accessible
on
this
this
resource.
B
B
Sorry,
the
running
services
that
are
on
this,
this
particular
node,
are
the
account
manager,
the
cloud
manager,
the
billing
manager,
the
slurm
controller
and
an
nfs
server,
so
that
this
management
node
is
is,
is
running.
This
service
stack
and
it's
it's
responsible
for
kind
of
configuring.
All
of
these
resources
and
managing
the
the
billing
and
and
interfacing
with
the
slurm
client,
that
is
on
the
login
node.
B
So
that
management
node
has
this
system
level
python
software
package
running
on
it,
and
this
service
is
to
interface
the
this
this
slurm
scheduler,
that
is
on
skyway,
which
again
this
is
a
separate
slurm
controller,
then
is
on
the
hpc
resource
and
the
this
this.
This
python
software
package
interfaces
the
storm
schedule
with
the
various
cloud
sdks.
B
So
requirements
for
running
this
software
stack
are
that
the
the
node
itself
has
an
nfs
server
running
and
it
has
slurm
installed
and
configured
the
python.
3
requirements
are
related
to
implementing
this
cloud
interface,
and
that
is
boto3
for
for
aws,
lib
cloud
for
the
other,
other
cloud
providers
and
pi
sql
for
managing
the
accounting
database
of
the
sorry.
The
billing
database.
B
The
billing
manager
is,
is
a
name
and
process
that
effectively
monitors
the
real-time
budget
and
the
cost
status
for
each
billing
account
and
the
cloud
manager
is
sort
of
the
brain
of
of
the
of
the
of
the
management
node.
It's
it's
a
daemon
process.
That's
continually
checking
the
status
of
the
slurm
queue
and
the
cloud
account
usage
information,
and
so
it
integrates
with
the
the
cloud
sdks
and
it
its
its
job
is
to
operate
on
the
cloud
nodes
by
either
creating
or
removing
or
annihilating
those
resources.
B
And
that's
that's
what
interfaces
with
the
billing
manager
so
from
the
from
the
client
they
can
issue
commands
to
interact
with
some
controller
and
if
they
submit
a
a
particular
job
that
jobs
the
the
law.
Script
that
is
running
on
this
resource
will
then
first
you
know
check
with
the
billing
manager
to
understand
whether
or
not
something
can
be
scheduled
if
the
nodes
themselves
aren't
provisioned.
B
You
know
they're,
not
they're
they're,
not
in
existence
on
the
cloud
resource
they
they'll
appear,
as
in
a
drain
state
to
the
user
until
they've
been
provisioned,
so
the
users
themselves
can
query
the
the
database,
the
stormdb,
because
they're
slammed
to
be
also
running
on
this
management
resource
at
any
time.
To
query
past
that
statistics.
B
B
Here
it's
google,
along
with
the
particular
type
of
resource
and
once
once
the
user
has
submitted
a
particular
job
and
it's
identified
that
they
have
sufficient
budget
and
the
type
of
resource
that
they've
requested
is
is
one
that
is
available
to
them.
Then
the
job
will
be
provisioned
and
this
will
go
from
drain
to
running
so
this
luau
plug-in
script.
That
interfaces
with
the
the
that
interfaces
with
the
the
rest
of
the
skyway
daemon
processes.
B
So
this
the
job
submissions
are
sent
to
the
scheduler
they're
checked
by
this
luau
submission
subscript,
and
then
you
know
it
depending
on
the
billing
module,
whether
or
not
it's
identified.
If
there's
resources
sufficient
billing,
sorry
sufficient
budget
left
and
the
resources
are
either
in
existence
or
or
not
available,
it
will
spin
them
up
and
start
the
start,
the
job.
B
That
depends
on,
if
you
specify
the
amount
of
time
that
you're
going
to
use,
otherwise
it
will
default
to
whatever
the
default
time
limit
is,
which
is
which
is
infinite.
So
it's
it's!
A
good
idea
to
specify
a
time
limit,
so
the
cloud
manager
daemon
is:
it
runs
on
the
sky
range
management
and
it's
it's
it's
in
it's
a
background
process.
It's
continually
checking
whether
or
not
to
create
or
release
cloud
nodes.
So
this
this
is
using
the
apaches
libcloud
python
library
to
manage
creation,
annihilation
of
resources
other
than
than
aws.
B
B
So
some
resources
are
are
protected.
As
I
mentioned,
the
I
o
nodes
and
the
storage,
those
are
not
cannot
be
destroyed,
they
must
persist.
So
here
is
an
example
of
the
of
the
configuration
file
which
specifies
the
type
of
node
instances
that
are
available
for
a
particular
a
particular
a
project
account.
B
So
it
looks
like
I'm
running
out
of
time
here,
so
I
want
to
quickly
the
flow
chart
of
actions
that
the
jobs
and
the
resources
go
through.
Here
is
the
the
the
brains
of
this
cloud.
Daemon
are
to
assess
if
the
budget's
exceeded
and
if
it
is,
then
it's
going
to
send,
send
a
message
to
kill
all
jobs
and
release
the
nodes.
If
there's
an
if
it's
not
exceeded,
then
it
checks
for
a
new
if
a
new
node
is
needed.
B
If,
if
the
answer
is
yes
and
it
checks
to
see
if
the
cost
rate
will
be
exceeded,
and
if
the
answer
is
no,
then
it
creates
a
new
node
and
does
other
operations
in
terms
of
getting
getting
the
getting
the
job
launched.
So
here
the
is
an
example
of
the
skyway
budget
system.
Notification,
the
the
billing
mod,
the
billing
module
itself,
monitors
all
of
the
real-time
budget
balance
and
makes
these
notifications
to
the
user.
B
So
what
about
bursts
into
resources?
Cloud
bursting
generally
means
switching
workloads
to
public
cloud
resources
when
local
resources
are
not
available.
So
what
I've
described
thus
far
does
not
account,
for
you
know
this
idea
of,
like
bursting
bursting,
to
cloud
where
things
just
automatically
happen.
B
They
have
to
copy
their
data,
and
you
know
on
the
premise
from
the
on-premise
mounted
storage,
this
project
space
to
the
cloud-mounted
scratch
base,
and
then
they
have
to
submit
the
job
to
the
scheduler
and
at
the
completion
of
the
job,
then
they
have
to
copy
back
their
outputs
back
to
the
on-prem
storage.
This
workflow
is,
you
know
rather
inconvenient
for
the
users
and
it's
it's
it's
sort
of
a
barrier
to
adoption
of
the
service.
B
Okay.
So
the
current
usage
statistics
of
this
this
resource
there
are
about
90,
total
users
that
have
used
it
to
date.
There
are
eight
different
projects
served.
There
have
been
about
2
000
jobs
submitted
and
about
18
000
of
cloud
credits
has
been
consumed
between
aws
and
gcp,
and
the
primary
users
for
this
resource
have
been
folks
in
education
courses.
B
So
things
with
jupiter,
notebooks
and
machine
learning
projects,
so
things
that
are
pain,
points
for
this
hybrid
implementation
are
the
managing
of
multiple
software
stacks
I
mentioned
so
each
cloud
provider
instance
right
now
has
its
own
globally
accessible
software
stack
for
the
user.
That
could
be,
you
know
rather
cumbersome
if
you
start
to
add
in
more
software
moving
forward
for
the
users
that
data
migration
is
one
of
the
bigger
pain
points.
B
We
didn't
have
gpus,
that's
one
example
of
using
the
the
cloud
for
that
purpose:
small
project
budgets:
again
there
are
faculty
that
don't
have
enough
money
to
buy
dedicated
resources,
and
so
they
want
to
just
simply
have
something
they
can
use.
Once
in
a
while
and
there's
you
know,
sort
of
it's
meant
to
have
zero
learning
curve,
but
there
there's.
You
know
there
are
some
things
that
could
be
improved
to
make
it
better
for
the
users,
and
so
with
that
I
oops.