►
Description
October 1st, 2021
Moderator: Niall Gaffney
Presenters: Shawn McKee, Henry Neeman, and Scott Yockel
Bios:
Shawn McKee is a Research Scientist at the University of Michigan Physics Department and Founding Director of the Center for Network and Storage-Enabled Collaborative Computational Science (CNSECCS), part of the Michigan Institute for Computational Discovery and Engineering (MICDE). He also directs the ATLAS Great Lakes Tier-2 (AGLT2) Center for the ATLAS Experiment at the CERN Large Hadron Collider. Since 2001, Dr. McKee has been the Network Project Manager for US ATLAS, planning for and developing the necessary network environment to support the US ATLAS computing model. He also co-leads the USATLAS facilities and distributed computing area and the ATLAS Distributed Data Management area. Since June 2012 he has led the Open Science Grid (OSG) Network area. He is the Principal Investigator on the “NSF CC*DNI DIBBs: Multi-Institutional Open Storage Research InfraStructure (MI-OSiRIS)” project and co-Principal Investigator on the”NSFCIF21 DIBBs: EI: SLATE and the Mobility of Capability” and “CC* Integration: NetBASILISK: NETwork Border At Scale Integrating and Leveraging Individual Security Components” projects.
Dr. Henry Neeman is the Director of the OU Supercomputing Center for Education & Research, Executive Director for Research Computing, Associate Professor in the College of Engineering and Adjunct Associate Professor in the School of Computer Science at the University of Oklahoma. He and Dana Brunson have been appointed joint co-leads of the XSEDE Campus Engagement program, which includes the Campus Champions. He received his BS in computer science and his BA in statistics with a minor in mathematics from the State University of New York at Buffalo in 1987, his MS in CS from the University of Illinois at Urbana-Champaign in 1990 and his PhD in CS from UIUC in 1996. Prior to coming to OU, Dr. Neeman was a postdoctoral research associate at the National Center for Supercomputing Applications at UIUC, and before that served as a graduate research assistant both at NCSA and at the Center for Supercomputing Research & Development. In addition to his own teaching and research, Dr. Neeman has collaborated with dozens of research groups, applying High Performance Computing techniques in fields such as numerical weather prediction, bioinformatics and genomics, data mining, high energy physics, astronomy, nanotechnology, petroleum reservoir management, river basin modeling and engineering optimization. He serves as an ad hoc advisor to student researchers in many of these fields. Dr. Neeman's research interests include high performance computing, scientific computing, parallel and distributed computing and computer science education.
A
You
know,
working
with
hubble
space,
telescope
archive
and
doing
a
lot
of
preservation
around
data
and
and
how
we
do
all
of
that.
This
is
an
important
topic
that
often
sort
of
gets
a
little
overlooked,
because
people
think
it's
it's
a
simple
process,
so
just
we'll
sort
of
hopefully
have
maybe
you
know
10
15
minute
talks
from
folks
and
then
leave
some
time
for
questions
somewhere
towards
the
end.
A
He's
part
of
the
the
great
lakes
atlas
tier
2
center,
which
is
part
of
another
experiment
producing
littles
of
data.
I
would
say-
and
preservation
of
that
is
sort
of
important
as
well,
but
this
is
part
of
the
atlas
experiment
coming
out
of
the
large
hadrian
collider
at
cern
since
2001
he's
been
the
network
project
manager
for
us
atlas,
making
planning
for
the
development
necessary
network
environments
to
support
the
computing
model.
He
also
co-leads
the
us
atlas
facilities
in
distributed
computing
area
and
the
atlas
distributed
data
management
area.
B
B
So
you
know
in
in
the
project
that
was
mentioned
osiris
we
tried
to
put
together
a
multi-institutional
multi-science
domain
storage
infrastructure
to
let
people
share
and
directly
access
data,
and
so
that
project
just
ended
at
the
end
of
august,
we're
continuing
to
operate
it
with
help
from
our
institutions,
and
so
this
sort
of
touches
on
one
of
the
one
of
the
challenges.
One
of
the
aspects
that
we
have
to
worry
about,
which
is
how
do
you
continue
to
provide
capability
services?
B
For
you
know
the
foreseeable
future
data
isn't
going
to
go
away.
It's
going
to
keep
getting
larger,
it's
going
to
keep
presenting
us
with
significant
challenges
and
there's
all
these
questions
around
data
provenance
data
life
cycle
management,
how
how
you
actually
keep
track
of
all
the
data
that
you've
generated,
how
you
reliably
store
it
and
provide
access
for
it
into
the
future,
and
you
know
I
have
to
say
that
I
obviously
don't
have
any
answers
here.
I've
seen
a
lot
of
the
challenges.
B
I'm
I'm
constantly
thinking
about
how
we
might
do
this
on
a
large
scale
for
the
atlas
experiment.
We
have,
as
mentioned
huge
amounts
of
data
we'll
be
producing
by
the
time
of
high
luminosity
lhc
half
an
exabyte
of
data
a
year
just
for
atlas,
and
that's
just
the
data.
Then
you
have
to
have
the
simulations
and
everything
that
goes
with
it,
and
where
do
you
put
that
indefinitely?
B
B
And
you
know
that
works,
but
you
know,
there's
a
cost,
and
so
these
labs
have
to
commit
to
each
one
being
a
steward.
For
you
know,
cern
has
a
complete
copy
for
each
experiment,
all
the
data
and
then
these
other
distributed
tier
1,
centers
national
scale
centers
around
the
world.
Each
take
roughly
10
percent
of
the
data
and
they're
responsible
for
that,
and
what
responsible
means
is,
of
course,
keeping
a
copy
for
the
future,
but
also
if
data
has
to
be
re-processed.
B
New
calibrations
come
out
new
methods
and
the
data
has
to
be
converted
back
into
something
new.
Then
that's
the
responsibility
of
each
one
of
those
sites,
so
that's
part
of
being
a
steward
of
the
data
is
if
there
is
actually
still
active
use
of
the
data,
somebody
has
to
provide
the
ability
to
actually
process
on
it.
Transform
it.
You
know,
apply
new
algorithms
to
the
data
based
on
new
discoveries
in
you
know
how
we
do
reconstruction,
how
we
do
calibration
et
cetera.
B
A
A
So
next
up
is
dr
henry
niemann,
who
is
the
director
of
the
oklahoma
university
super
computing
center
for
education
and
research
and
executive
director
for
research
computing
associate
professor
in
the
college
of
engineering
and
adjunct
professor
in
the
school
of
computer
science?
All
at
the
university
of
oklahoma,
he
and
dana
brunson
have
been
appointed
co-lead
to
the
exceed
campus
engagement
program,
which
includes
mathematics
from
the
state.
Sorry,
which
includes
the
campus
champions,
which
is
a
great
program
that
really
helps
bring
cyber
infrastructure
across
to
many
people
across
the
country.
A
He
received
his
bachelor
of
science
and
ba
in
statistics
with
a
minor
in
mathematics
from
suny
buffalo
in
1987,
his
master
science
in
computer
science,
from
the
university
of
illinois
urbana-champaign
and
phd
in
1996
prior
to
ou.
He
was
a
postdoctoral
researcher
at
the
national
center
for
supercomputing
application
at
illinois
and
served
as
a
graduate
research
assistant,
both
ncsi
and
for
the
center
for
super
computing
research
and
development.
A
In
addition
to
his
own
teaching
and
research,
he's
collaborated
with
many
different
research
groups,
applying
high
performance
computing
techniques
in
fields
such
as
numerical
weather,
predictions
bioinformatics,
genomes,
genomics
data
mining,
high
energy
physics,
astronomy,
nanotechnology.
The
list
goes
on
and
on,
and
he
serves
as
an
ad
hoc
advisor
for
student
research
in
many
of
these
fields.
A
His
research
interests
include
high
performance,
computing,
scientific
computing
and
distributed
parallel
computing,
but
I
also
know
henry
is
involved
in
the
our
store
project,
which
is
a
long-term
data
preservation
project
funded
by
the
nsf.
So
I
will
turn
it
over
to
henry
to
tell
us
about
data
preservation.
From
his
point
of
view,.
C
A
C
All
right,
that's
just
a
title
slide.
So
don't
worry
about
that
at
ou
we
are
so
dumb.
We
are
deploying
five
major
systems
at
the
same
time,
so
I
just
wanted
to
point
that
out.
So
any
slowdowns
in
getting
stuff
out
to
you
all
is
because
we're
trying
to
do
many,
many
things
on
very
limited
time,
so
the
project
we're
doing
at
ou
is
called
the
ou
and
regional
research
store.
It's
basically
a
giant,
robotic
tape
library.
So
it
was
delightful
to
hear
sean
say
that
tape
is
really
important.
C
Tape
is
not
important,
because
tape
is
good.
In
fact,
tape
sucks,
but
tape
sucks
the
way,
democracy
and
old
age
suck
it
sucks,
except
compared
to
the
alternative,
because
the
alternative
to
tape
is
not
disk.
The
alternative
to
tape
is
throwing
the
data
away
and
there's
a
very
simple
reason
for
this.
C
The
business
model
is
a
business
model
that
makes
archiving
affordable,
but
the
problem
with
archiving
is
that
archiving
is
expensive,
but
in
the
sense
that
there
are
giant
fixed
costs,
tape,
systems,
the
hardware
and
the
software
are
terribly
expensive
and
they're
terribly
expensive.
Before
you
put
the
first
slot
for
a
tape
cartridge
on
the
floor,
we
found
that
over
80
percent
of
the
hardware
and
software
cost
we
had
was
fixed
costs
that
came
with
730
tape,
cartridge
slots
and
then
adding
more
slots
after
that
was
super
cheap.
C
So
we
have
an
nsf
grant
that
pays
for
the
hardware
and
software
and
the
initial
warranties.
Our
cio
covers
the
mechanical
space
power,
cooling
network
and
then
also
the
labor
as
part
of
the
cost
share,
and
then
he
also
picks
up
the
the
annual
support
after
the
initial
warranties
expire.
So
what
do
researchers
pay?
They
just
pay
the
incremental
cost,
which
is
the
tape
cartridges.
So
researchers
buy
tape
cartridges.
They
don't
buy
anything
from
us.
We
never
touch
their
money.
They
have
the
tape
cartridges
shipped
to
us.
C
We
pop
them
in
the
tape
archive,
and
that
is
super
cheap,
so
lto.
Actually,
when
lto9
was
released
a
couple
of
weeks
ago,
lto7
tape
cartridges
came
down
by
about
10
percent
right
around
that
same
time,
so
they're
now
about
six
bucks
per
per
usable,
terabyte
per
copy
or
12
bucks
for
dual
copies,
which
is
usually
how
we
present
it
to
people-
and
you
know,
tape
lasts
a
long
time.
C
So
when
you
stick
it
in
the
tape
archive
it's
there
for
the
life
of
the
archive,
now
sean's
point
about
later,
moving
it
to
better
more
recent
versions
of
the
of
the
tape.
That
is
a
point
well
taken,
but
we're
not
doing
that
all
the
time
we
do
it
once
in
a
great
while
in
fact
we're
just
about
on
monday,
we
will
start
doing
it
here,
copying
from
lto,
5
and
6
to
lto7
type
m.
C
We're
starting
with
six
drives
gives
us
a
little
under
two
gigabytes
per
second
aggregate
bandwidth,
but
that's
pretty
good
for
what
people
need
and
then
we're
going
to
add.
Some
more
tape
drives
of
later
flavors
later
on,
depending
on
how
many
no
cost
extensions,
we
can
get
improv
approved
on
our
nsf
grant.
We've
got
about
half
a
petabyte
of
disk
as
a
front-end
landing
pad
and
we
are
making
sure
that
the
secondary
copies
live
on
the
other
campus,
not
in
the
tape
library.
C
That
would
be
a
complete
waste
of
tape,
cartridge
slots,
but
on
shelving
units
in
an
environmentally
controlled
room
so
that
we
are
actually
doubling
not
just
the
number
of
slots,
but
the
actual
capacity
for
content
by
putting
in
the
second
tape
library,
keeping
the
secondary
copy
inside
the
tape.
Library
to
us
looks
like
a
waste
of
money
and
then
what
kind
of
features
are
we
are
we
adding
that
we
didn't
used
to
have
on
our
old
tape
archive
one
thing
we're
doing
we
haven't
written
the
script.
C
Yet,
although
we
have
sort
of
a
design
for
the
process
is
auto
archiving.
So
when
a
user
drops
a
file
into
a
particular
directory,
a
daemon
script
comes
along
notices.
It's
there
and
automatically
archives
it
to
tape.
We
could
have
used
expensive
commercial
software
to
do
that,
but
it
would
have
been
expensive
and
we're
always
broke
so
we're
doing
that
by
hand.
C
File
sharing
we've
got
a
globus
license
that
we
just
got
actually
a
temporary
license
until
we
get
the
paperwork
worked
out
because
we're
doing
the
hipaa
version
with
the
baa.
That's
got
to
go
through
legal.
You
know
how
it
is,
but
so
we've
got
globus
file
sharing
is
something
we'll
have
in
production
in
the
next
few
weeks.
C
If
a
user
is
really
itching
to
make
sure
that
a
file
of
theirs
is
always
on
tape,
they
can
buy
disk
we're
just
not
recommending
that
we
don't
expect
a
lot
of
uptake,
because
it's
a
whole
lot
of
money
and
honestly
not
that
much
value
and
then
the
price
I
mentioned.
The
price
per
terabyte
has
come
way
way
down
and
I
think
we'll
continue
to
fall.
C
What
we
have
seen
historically
is
that
the
most
recent
version
of
lto
costs
roughly
twice
as
much
per
terabyte,
so
we
never
ever
recommend
the
most
recent
version
of
lgo,
even
though
it
would
save
us
on
tape
perforated
slots,
because
our
users
would
be
pissing
their
money
away
and
then
they
would
be
mad
and
stock
off
in
a
huff,
and
we
don't
need
that.
I
think
that's
pretty
much
it.
I
do
have
a
slide
on
on
our
disc
stuff,
but
that's
not
really
relevant
here.
C
A
C
C
A
Sure,
absolutely
all
right
and
let
me
introduce
our
third
presenter
today.
Our
third
panelist
is
scott
yoko
who
joins
us
from
harvard
where
he's
been
leading
the
hbc
group
in
science
and
engineering
since
2015
and
is
working
now
in
the
university
of
research,
compu
computing
office
he's
been
on
both
sides
of
hpc,
both
as
a
researcher
and
administrator.
So
he
has
that
understanding
of
of
of
you
know
what
it
is
to
actually
work
with
data
and
work
with
modeling.
A
His
research
domains
include
computational
chemistry,
with
overlaps
into
material
science.
Mechanical
engineering
he's
involved
also
in
the
campus
champions
project
and
is
co-pi
for
the
northeast
storage
exchange,
the
new
england,
research
cloud
and
the
ci
center
for
excellence,
creating
a
national
research
computing
data
center
and
resource
center.
So
why
don't
you
jump
on
in
and
let
us
know
about
sort
of
the
aspects
of
data
preservation.
You
have
it
on
your
hands
sure
I.
D
So
I'm
going
to
preface
that
it's
actually
my
first
time
to
be
on
campus
in
an
office
in
18
plus
months
and
then
and
in
30
minutes.
I'm
having
happy
hour
with
my
group
for
the
first
time
in
like
20
months,
so
anyways
today
across
harvard
there's
a
very
disparate
kind
of
understanding
of
data
and
usage.
And
what
that
is
is
a
problem.
D
I've
learned
a
lot
through
collaborating
with
merced
crosses
over
the
years
who's,
the
kind
of
lead
pi
on
harvard
on
dataverse.org,
but
also
helps
run
the
harvard
dataverse.
She
just
recently
took
a
position
this
summer
back
into
she's,
actually
a
secretary
of
open
data
in
the
new
catalonian
government
anyways.
So
I've
learned
a
lot
in
that
space
and
what's
happened,
is
through
kind
of
an
evolution.
D
There
has
been
a
much
tighter
knit
thing,
that's
happening
between
research,
computing
and
our
library
and
that's
kind
of
what
I'm
going
to
talk
about
a
little
bit
today,
so
a
few
slides
in
that
space,
so
one
is
around
the
problem.
So
typically
the
the
friction
happens
when
researchers
would
like
to
collaborate
from
different
institutions
right,
it's
not
easy.
D
I
grew
up
in
the
sciences
and
this
and
I
was
a
theoretician
for
the
most
part.
So
this
to
me
didn't
seem
like
a
big
deal
until
I
started
engaging
a
lot
with
more
social
scientists
and
stuff
and
so
like,
where
they
pull
their
data.
What
do
they
call
their
data
set
is
a
whole
different
idea
right.
It's
a
much
different
refinement
until
you
get
to
the
data
set
that
you
would
actively
do,
research
on
which
becomes
the
analytical
data
set
right,
because
this
data
is
all
messy
right
for
most
of
the
scientists.
D
D
So
inside
of
an
individual
institution,
we
may
be
able
to
have
a
lot
more
control
right.
Around
data
collection
have
a
data
use
agreement
system
that
the
contracts
go
through
understand
who
has
access
to
that
data?
We
can
have
some
kind
of
interoperability
space.
That
is
say
the
research
computing
cluster
or
some
specialized
secure
enclave
that
we
onboard
multiple
different.
D
You
know,
groups
to
this
thing
in
the
middle
could
exist
on
our
campus.
It
could
exist
as
a
shared
infrastructure,
a
cloud
resource
or
something
like
that
right,
but
I
think
the
thing
that
that's
still
a
friction
even
on
our
internal
campuses
is
that
the
tooling
that
we
have
to
do
these
is
all
separate,
they're,
all
normally
different
pieces
of
technology
and
into
a
researcher.
D
This
creates
a
lot
of
friction
in
their
process
and
they
get
really
frustrated,
and
I
think
this
is
where
we
have
not
collectively
done
as
good
a
job
as
we
should
and
then
look
at
other
tools
that
they
use,
like
google
docs
right
great
way
to
collaborate.
Google
drive
super
simple
for
them
to
engage
with.
We
don't
have
the
analogous
of
those
on
our
large
campus
infrastructure
and
I
think
that's
the
kind
of
next
tranche
of
work
to
do
and
not
think
about
them.
D
As
an
individual
thing
like
we
could
put
open
on
demand
in
front
of
our
hpc
system.
That's
great,
but
if
it's
not
talking
to
the
data
publication
space,
if
it
doesn't
understand
the
doa
system,
then
all
these
things
become
kind
of
very
manual
labor-intensive,
different
communities
of
people
that
a
faculty
has
to
engage
with,
and
that's
that's
kind
of
the
problem
currently.
D
So
if
you
were
then
going
to
to
have
them
do
that
same
type
of
thing
at
a
different
institution,
there
has
to
be
some
kind
of
sets
of
standards
to
to
do
the
same
kind
of
tooling
across
multiple
sites
and
that's
when
you
get
kind
of
into
a
federation
of
services
and
that's
where
kind
of
a
true
global
solution
of
data
commons,
and
there
has
been
people
that
have
thought
about
those
types
of
things
like
the
nih
spent
quite
a
bit
of
time.
Thinking
about
that
across
their
various
different
institutes.
D
D
We
thought
about
what
are
all
the
other
projects
that
are
out
there
and
people
what
are
their
concerns
in
the
state
of
common
space,
and
so
at
harvard
for
more
than
10
years.
We've
had
this
project
dataverse
there's
a
local
installation
that
is
harvard
dataverse
and
it
is
the
largest
of
the
different.
I
think
there
are
like
120
different
dataverse
installations
worldwide,
so
it
has
a
pretty
extensive
data
size
of
data
set.
It
has
quite
a
bit
of
downloads.
This
is
also
like
a
year
old
stats.
D
So
that's
it's
a
little
bit
more
than
that
now,
but
there
is
some
limitations
into
the
code
base.
It's
set
up.
It
was
designed
for
social
science
stuff.
There
are
some
places
where
it
kind
of
is
challenging
for
large
scientific
data
sets,
and
that's
one
of
the
things
we're
trying
to
overcome,
because
the
mechanism
to
move
data
currently
is
through
just
through
a
web
or
a
web
api.
D
D
The
only
place
that
harvard
dataverse
is
engaging
in.
Is
this
last
part.
So
it's
an
afterthought
in
the
process
of
research
that
somebody
thinks
about
how
should
I
tag
my
data?
How
should
I
represent
versions
of
this
data
along
the
way
in
the
data
analysis
phase?
So
what
we'd
like
to
do
is
create
a
set
of
tooling
so
that
they
can
check
in
these
raw
unpublished
data
sets
right.
They
if
needed,
they
can
get
a
doi,
but
there
at
least
is
metadata.
That's
represented
to
them.
D
This
metadata
is
really
really
important
because
it
would
represent
the
state
of
the
data,
but
we
can
also
include
things
like
the
code,
the
container
or
other
artifacts
that
are
part
of
the
data
analysis
process,
like
maybe
at
the
data
workflow
like
using.
You,
know,
dag
and
other
common
workflow
language,
those
types
of
things
to
check
in
with
the
data
source,
so
that
rich
metadata
now
represents
really
the
providence.
We
can
then
now
use
this
metadata
with
a
variety
of
different
tools
and
things
so
on.
D
The
left
side
is
trying
to
represent
what
kinds
of
tools
that
exist
in
the
data
analysis
space
that
we
want
to
integrate
with.
So
our
space
is
one
that
we
we're
trying
to
integrate,
dataverse
with,
which
is
an
electronic
lab
notebook.
So
that's
the
space
where
they're
explaining
what
they're
doing
right.
It's
not
a
place
in
which
the
computations
are
happening.
D
D
We
can
capture
those
in
a
repository
like
like
dataverse,
so
they
already
talked
about
globus.
We
know
globalists
can
be
used
as
a
you
know,
data
management
system
it
can
also
be
used
for
sharing
but
kind
of
the
thing
that's
lacking.
Is
you
don't
go
to
the
globus
interface
and
search
for
data
right?
You
can
find
an
endpoint.
D
You
can
find
a
shared
endpoint
or
a
management
point,
and
so
to
be
able
to
integrate
the
metadata
part
that
we
have
in
something
like
dataverse,
so
that
you
can
find
the
data
and
then,
instead
of
it
being
represented
as
a
zip
file
that
you're
going
to
download
via
the
web,
you
could
actually
have
automatic
created
shares
that
happen
in
globus.
So
that's
one
of
the
other
integration
points,
that's
underway.
D
First
kind
of
starfish
user
group
meeting,
which
was
I
actually
it
was
information
that
I
had
talked
to
other
users
and
we
decided
to
do
is
ourselves
as
a
grassroots
effort.
I
would
say
that
there
is
a
vibrant,
rich
community
among
various
different
universities,
as
well
as
a
lot
of
kind
of
biotech
people
where
you
are
starting
to
be
able
to
do
some
pretty
interesting
flows
of
data
by
tagging
or
creating
policy.
D
D
That
is
not
the
command
line,
interface
to
run
bu
and
wait
till
tomorrow
to
see
who's
using
what
they'll
have
instantaneous
access
from
a
database
about
data
in
a
much
grander
scheme
with
visualization.
So
you
could
choose
lots
of
different
solutions
here,
but
this
is
meant
just
to
represent.
You
need
something
like
that:
that's
not
a
command
line
interface,
because
there
are
so
much
better
tools
in
the
cloud
space
they
become
accustomed
to.
We
have
to
provide
them
that
they
also
just
should
be
using
command
line
to
move
large
amounts
of
data
around
anymore
right.
D
So
that's
kind
of
our
idea
and
I'll
spend
a
couple
more
slides.
Talking
about
this.
What
is
the
storage
tier
three
space,
because
this
is
something
new
that
we
didn't
have
before
tier
one?
Is,
you
know
like
adjacent
to
high
performance
computing?
Tier
two
would
be
like
kind
of
your
generic
nsf
file
system.
D
So
a
couple
more
slides
to
talk
about
the
tier
three.
So
as
a
part
of
both
that
tier,
two
and
tier
three,
there
is
the
new
england
storage
change
project,
which
is
one
of
the
divs
project,
and
my
group
has
been
kind
of
lead
at
running
the
back
end
engineering
on
it.
It's
a
collaboration
across
all
the
schools,
with
dynamic,
hpc
and
mghbc
as
as
an
entity
itself.
So
here's
the
copyis
and
to
just
give
you
two
diagrams
here
to
finish
off
what
we've
been
doing
as
so.
D
So
it
was
either
bite
the
bullet
around
our
own
local
federation
identity,
where
every
a.d
across
not
just
a
single
right
university,
but
it
could
be
across
lots
of
different
parts
of
the
university.
We
have
to
tie
into
some
centralized
identity
management
for
nghbcc,
in
which
there
are
at
least
20
to
30
000,
individual
users
across
right
now
on
our
systems
across
different
schools.
That
didn't
seem
like
a
good
idea
to
try
to
do
so.
D
What
we
did
is
we
completely
abstracted
the
idea
of
of
the
user
identity
from
the
data,
so
data
is
represented
in
ceph.
It
can
be
represented
through
the
data
transfer
node
via
cfs
and
on
the
tape
system
that
we
just
employed.
So
this
is
a
ibm
ts,
4500,
it's
about
130
petabytes
of
capacity
with
one
petabyte
of
spectrum
screen
on
the
front
end
anyway.
So
that's
that's
also
mapped
up
here.
D
The
data
transfer
nodes
so
what's
interesting,
is
that
the
it
turns
out
the
only
people
that
really
need
access
to
the
data
transfer
nodes
are
not
the
actual
end
users,
not
the
researchers,
it's
the
tape
owner
or
the
media
owner
of
of
ceph
that
have
added
stuff.
So
this
is
the
other
research
computing
admins
right.
So
then
see
engineers,
which
is
really
one
person
engages
with
this
team
of
like
15
or
so
tops,
and
they
can
log
in
directly
here
and
they
can
create
a
global
share.
D
You
know
that
they
can
see
as
the
managed
endpoint
then
in
the
global
sense
they
can.
They
can
then
share
that
out
to
any
globus
id,
so
they
can
add
the
pi
as
the
data
manager
and
then
they
can
do
what
they
want
with
it.
So
in
that
point
it
becomes
their
own
internal
research,
computing
group
say
from
boston
university
and
their
community
to
manage
how
they
want
to
do
this
process.
D
So
it's
not
meant
to
replace,
like
your
high
performance
computing
attached
storage,
it's
meant
to
be
like
that
next
step,
that
kind
of
large-scale,
cheaper
staging
storage
and
so
relative
cost
wise.
The
tape
stuff
is
about
five
dollars
per
terabyte
to
operate,
and
the
self
storage
is
on
the
order
of
like
40
dollars
per
terabyte
per
year
to
operate,
and
then
your
high
performance
stuff
might
be
a
little
bit
more.
Maybe
it's
like
a
hundred
dollars
per
terabyte
per
year,
depending
on
what
you're
doing,
if
you
have
any
other
costs.
D
So
what's
interesting
is
that
this
solves
it
for
that
kind
of
posix
use
case,
but
then
whoa
it
dropped
out
of
the
sorry
about
that.
I
clicked
on
the
diagram
and
then
it
went
to
the.
But
if
you
wanted
to
create
an
s3
bucket,
so
you
have
like
global
name
space.
There's
a
cool
piece
of
code
called
min
io.
You
can
run
it
in
a
very
similar
fashion,
so
we
have
cfs
and
we
have
gps
on
the
back
end.
D
And
so
when
we're
mapping
like
on
this
instance
like
we
have
like
a
tape
pool,
we
map
that
to
the
data
transfer
node
with
a
file
set
which
has
we
can
map
that
then
to
minio
and
either
two
mechanisms.
We
can
create
a
min
io
tenant
space,
which
is
kind
of
a
name
space
that
then
this
person
here,
the
tape
owner
or
the
rc
provider.
They're
like
they
have
an
admin
interface
and
then
they
can
create
buckets.
D
However,
they
want
and
they
can
share
one
large
tape,
pool
or
one
large
disk
pool
and
share
out
buckets
and
access
keys.
So
what
we've
done
is
completely
separate
the
user
from
the
data
as
far
as
the
management
of
it
and
what
happens
is
they
have
access
tokens?
So
it's
at
this
access
layer
is
where
the
id
the
identity
of
users
exists,
and
so
that's
kind
of
been
an
interesting
thing
that
we've
done
and
one
of
the
places
that's
been
nice.
D
Is
that,
like
in
the
harvard
library
they
have
a
digital
repository
service
which
is
like
part
of
their
archive
service
and
their
requirements
to
do
archiving
are
two
disk
copies
and
two
tape
copies,
so
they
were
managing
all
the
infrastructure
themselves
and
as
they've
transitioned
onto
different
services,
they're
now
using
s3
as
the
mechanism
to
store
data
and
transfer
data
and
access
data,
so
they're
able
to
leverage
the
min.
I
o
system
here
have
a
copy
on
tape.
D
So
that's
just
kind
of
an
example
of
of
things
that
we
were
interested
in
things
that
are,
we
find
a
problem
and
not
necessarily
like
100
solution,
but
we
do
have
some
building
blocks.
I'd
like
to
think
that,
especially
like
from
this,
we
are
working
on
various
different
components
of
this
and
building
blocks,
we're
hoping
to
get
to
the
point
where
we
have
kind
of
some
standardization
across
this
kind
of
api
framework.