►
From YouTube: ESPA Module 3C - Hardware Requirements
Description
Joe Hoang (Software Engineer) with PiKNiK discusses hardware requirements at the Enterprise Storage Provider Accelerator (ESPA) bootcamp week that took place in February 2022. This discussion includes environmental variables, hardware architecture, examples, what scales and what doesn’t, deal and data flow, racks and power.
Data is growing at an incredible speed and much of this data is archived and/or simply lost by enterprises. Our program will accelerate net-new Enterprise Storage Providers into the ecosystem, using a Web3 protocol with an impressive incentive model called Filecoin.
Learn more at:
Sign-up: https://web3espa.io
Landing Page: https://m.fil.org/espa-bootcamp
Follow ESPA:
Twitter: https://twitter.com/web3ESPA
LinkedIn: https://www.linkedin.com/company/web3espa/
A
All
right,
everybody
how
you
doing
today,
good
I
hope
yeah.
So
I
know
it's
a
little
jarring
to
see
my
name
software
engineer
and
now
I'm
talking
about
hardware
requirements,
but
I
do
have
some
experience
in
hardware
field.
Just
you
know
as
a
kid
playing
around
with
computer
parts,
but
also
when
I
first
got
to
picnic
and
even
to
this
day
I
still
do
a
good
amount
of
system,
administrative
work.
A
A
Okay
before
we
get
to
that,
though,
I
do
want
to
say
that
some
of
the
numbers
that
we
use
in
some
of
our
systems
are
based
on
environmental
variables,
basically,
values
that
you
can
set,
and
then
those
can
change
the
way
some
processes
run
and
then
also
how
some
hardware
pieces
are
ran
and
I'll
give
an
example
of
that
later,
as
well
all
right.
So
first,
let's
talk
about
hardware
architecture,
things
you
want
to
consider
when
you're
planning
out
your
servers.
A
What
hardware
pieces
you
want
and
really
the
main
thing
you
want
to
consider-
are
the
responsibilities
of
what
you
want.
The
servers
to
do
basically,
the
lotus
components
that
alex
was
going
over
earlier.
So
I
went
ahead
and
created
a
chart
here
of
the
six
main
responsibilities.
The
three
on
the
left,
pc1,
precurement1,
pc2,
pre-commit,
2
and
c2-
commit
2
are
the
ones
that
are
in
charge
of
the
sealing
process
actually
getting
that
sector
through
the
ceiling
pipeline.
A
Those
are
normally
referred
to
as
your
worker
nodes.
Those
are
the
ones
just
going
through
the
processes,
the
seal
and
then
the
ones
on
the
right.
The
lotus
miner
node,
the
lowest
damage
node
and
the
lotus
market
node
are
the
ones
that
those
your
storage
provider
needs
your
deals
and
also
connecting
to
the
blockchain.
A
A
If
there
is
a
check
mark,
that
means
it
is
a
process
that
does
use
that
hardware
piece
intensively.
For
example,
if
you
take
pc1,
it
requires
64
gigabytes
per
second,
it
works
on
so
there's
a
check
mark
there.
If
there
is
an
x
without
an
explanation
afterwards,
that
means
it
does
not
require
that
piece
of
equipment
at
all.
So
using
the
same
process
pc1,
you
can
see
that
it
does
not
require
any
gpu
usage
if
there's
an
x,
but
there
is
an
explanation
afterwards.
A
That
means
that,
while
it's
not
required
by
the
the
process
specific
or
yeah,
it's
not
required
by
the
process,
it
uses
it
very
non-intensively.
So
for
the
pc2
you
see
that
it
uses
about
15
or
16
gigabytes
per
sector,
which
may
be
a
lot
if
you
have
double
computers
before.
But
if
you
compare
to
the
two
next
to
it,
pc1s
and
c2s,
you
can
see
that
it
is
relatively
low
comparatively
and
then.
A
Okay,
hopefully,
you've
had
enough
time
to
look
at
that.
So
then
some
general
recommendations
when
it
comes
to
the
hardware
that
you
use
the
first
thing
that
we've
mentioned
many
times
is
that
you
only
want
to
use
enterprise
level
equipment
and
you
want
to
avoid
consumer
grade
equipment.
Enterprise
level
equipment
are
meant
to
last
longer
and
they're
meant
to
run
24
7,
which
is
usually
the
case
when
you're
running
an
enterprise
level
storage
provider.
A
It
may
seem
very
appealing
at
first
when
you're
online,
looking
for
your
parts,
whether
it
become
price
convenience.
Oh
look,
there's
someone
selling
on
ebay,
but
we
I
can
promise
you
that
cutting
corners
will
end
up
in
a
bad
result
for
you
once
the
honeymoon
phase
is
over
and
the
piece
of
equipment
is
broken
now
you're
sad
and
have
to
go,
buy
another
one.
A
Getting
a
little
more
specific,
I
want
to
mention
some
scratch
space,
so
the
scratch
space
that
you
use
for
sealing
on
your
pc
ones
and
pc2s
for
those
you
typically
want
to
avoid
spinning
discs.
They
won't
be
performing
enough
for
the
reed,
intensive
and
right
intensive
processes
and
then
again
consumer
drives.
Won't
last
very
long,
if
you
remember
alex's
anecdote
about
his
sabran
that
went
that
reached
over
100
usage
in
a
few
months.
That's
a
good
example
of
why
you
shouldn't
use
one
cpus
in
general.
A
More
cpu
cores
just
equals
better
performance,
it's
just
a
numbers
game,
but
you
also
want
to
make
sure
that
you
don't
over
provision
your
servers
because
now
you're
going
to
have
extra
resources
that
I
can't
use
properly
and
then
also
for
the
cpus
on
the
documents
they
mention.
How
models
with
intel
shaw
extensions
will
work
better.
A
That's
because
they
do
use
the
shaw
check
sum
within
the
proof
and
then
lastly,
gpus
similar
cpus,
more
cores
equals
better
performance,
and
you
also
want
to
be
bell
person
compatible
where
a
list
can
be
found
in
the
internship
documents.
Okay,
so
now
we've
gone
over
just
the
general
recommendations.
I'm
also
going
to
talk
about
some
examples.
How
picnic
does
it
specifically
before
I
get
into
that?
A
I
do
want
to
mention
that
I've
been
using
two
terms
that
may
seem
interchangeable
workers
and
servers,
but
I
want
to
differentiate
the
fact
that
a
worker
is
just
referring
to
the
process,
while
the
server
might
be
more
akin
to
what
you
think
about
for,
like
an
actual
computer,
the
hardware
pieces,
the
things
you
plug
in
then
you
rack,
okay,
so
first
one
is
going
to
be
the
pc
one
server
on
this
one.
I
the
first
thing
I
want
to
highlight
is
the
64
core
cpu.
A
If
you
remember
me
mentioning
environmental
variables,
one
of
the
ones
that
we
do
use
is
called
multi-core
sdr.
What
that
does
is
that
it
will
enables
the
cpu
to
use
a
whole
core
complex,
rather
than
just
a
single
core
to
work
on
a
cpu.
This
will
decrease
the
ceiling
time
so
doing
that
it
will
use
four
cores
per
sector
and
then,
with
a
64
core
cpu,
we
can
go
up
to
16
sectors
in
parallel
or
15.
A
If
you
don't
want
to
max
out
your
cpu
completely
off
of
that
is
how
we
derive
our
ram
and
our
scratch
space.
The
ram
at
1024
and
then
uses
64
gibbs
per
sector
calculates
to
about
15
16.
and
then.
Similarly,
if
you
remember
our
talk
about
proof
of
replication,
where
the
data
in
a
sector
is
replicated
many
times
to
reach
about
480
gigabytes
with
a
7.6
tib
u2
drive,
it
also
calculates
out
to
about
15
and
16.
A
Next
is
our
pc2
server
that
will
do
pc2
work.
The
core
part
for
this
server
would
be
the
gpus,
with
three
gpus
powerful
enough
to
host
two
workers
on
each
one
or
two
worker
processes.
You
can
go
ahead
and
fit
six
pc2
workers
on
this
pc2
server,
with
six
workers
being
able
to
work
on
six
sectors
in
parallel
and
about
15
to
16
gigabytes
per
that
comes
out
to
about
90,
to
96
gigabytes
required
for
the
processes.
A
A
A
And
for
the
last
worker
server
that
we
have
is
the
c2
server.
Similarly,
it
has
four
four
gpus
instead
of
three,
but
it
also
fits
two
workers
per
gpu,
giving
us
eight
workers
on
a
single
server
at
eight
workers
on
a
single
server
and
192
gigabytes
used
per
worker
that
all
that
equates
to
about
1
536
gigabytes
of
ram.
A
Okay,
so
that's
the
end
of
the
sealing
process
or
the
ceiling
pipeline
workers.
Now
we're
going
to
go
into
more
of
the
storage
provider
nodes
and
first
probably
most
important-
is
the
lotus
miner
node.
This
is
the
one
that's
going
to
be
handling
your
daily
audit
on
window
posts.
You
know,
failing
that,
you
get
faulted.
Failing
that
again,
you
get
slash
this
guy's
gonna,
be
your
most
important
one
go
ahead
and
give
it
256
gigabytes
of
ram,
because
at
the
highest
points
of
window
post
it
can
reach
up
to
250
gigabytes
of
ram
usage.
A
A
We
give
a
16
core
cpu
because
it
does
have
intensive
usage
at
the
beginning
and
end
of
its
proof,
just
like
the
c2
and
then
it
also
doesn't
need
scrap
space
because
it
doesn't
deal
with
any
sectors
long
term,
and
then
it
also
uses
a
gpu
to
do
its
calculations.
I
do
want
to
mention
that
the
calculations
can
be
done
by
the
cpu.
However,
if
you
decide
to
go
that
route,
the
cpu
will
generally
perform
worse
and
will
burn
out
quicker
than
a
gpu.
A
A
However,
picnic
does
host
many
different
minor
nodes,
so
instead
of
running
an
individual
lotus,
daemon
node
on
every
single
server
that
also
hosts
the
lotus
miner
node
we
go
ahead
and
just
create
an
individual
server
for
it
that
all
of
them
connect
to
for
this
one
we're
given
256
gigabytes
of
ram,
which
is
definitely
over
provisioning,
but
again,
like
I
mentioned
since
we
have
so
many
miners
connected
to
it.
The
last
thing
we
want
is
for
it
to
go
down,
and
now
every
single
one
of
the
miner
cannot
connect
to
the
blockchain.
A
Lastly,
is
the
market
node
this
one?
It
can
also
be
ran
on
the
miner
node.
If
you
would
like
to
to
provision
a
single
server
to
do
that,
but
here
at
picnic
we
like
to
individualize
that
as
well,
we
give
it
one
terabyte
of
ram.
This
is
for
the
sectors
that
have
to
import.
When
you
import
a
sector,
it
uses
32
to
64
gigabytes
of
ram,
and
then,
since
we
are
doing
a
lot
of
deals,
we
would
like
to
parallelize
it
to
make
sure
they
come
in
and
are
as
smoothly
as
possible.
A
A
Now
I
do
also
want
to
mention
that
these
are
just
the
examples
that
picnic
uses
and
how
we
do
it.
Obviously,
when
you
guys
create
your
workers
or
your
different
nodes,
you
may
find
that
a
pc2
worker
and
a
c2
worker
work
really
well
together.
You
may
find
a
way
to
combine
a
pc1
and
c2
worker,
and
then
you
create
a
hyper
efficient
server,
but
at
picnic
we
like
to
usually
stick
with
the.
I
believe
it's
called
the
one
to
one
bare
metal
servers.
A
So
now
we
can
talk
about
what
scales
and
what
doesn't
with
focus
on
space
and
electricity.
So
first,
let's
talk
about
some
some
rack
space.
Before
I
get
started,
I
want
to
mention
how
you
may
notice
a
unit.
That's
just
labeled
you!
If
you
are
familiar
with
data
centers,
then
you
already
know
that
that
stands
for
rack
unit.
If
not
a
rack
unit
is
just
the
standard
unit
used
in
racks
at
data
centers
to
measure
the
height
of
a
server.
A
So
if
you
recall,
a
c2
server
will
have
four
gpus
per
server
and
they'll
have
two
workers
per
gpu
and
then
with
11
servers
within
the
rack.
That
gives
you
88
total
workers
and
one
rack
after
some
googling
and
a
lot
of
averaging
and
guessing,
I
made
a
random
number
of
1.2
000
a
month
for
a
single
rack
at
a
data
center
and
then,
if
you
went
ahead
and
decided
condensing
is
too
much
work
or
it's
too
much
of
a
headache.
You
want
to
separate
every
single
worker.
A
Now
you
have
to
run,
or
now
you
have
to
rent
out
eight
times
the
amount
of
racks,
putting
your
costs
at
about
10k
a
month.
Now,
albeit
that
is
an
extreme
case,
you
probably
won't
be
doing
that.
But
what
you'll
more
realistically
find
is
that
some
server
cases
you
find
or
motherboards
that
you
may
find
aren't
able
to
support
four
entire
gpus
on
it.
It
may
only
handle
three,
maybe
two
but
you'll,
find
that
even
then
your
rackspace
cost
could
go
up
50
to
33
or
however
many
proportionately
to
what
you
take
out.
A
Well,
when
you
have
that
many
hardware
pieces
in
a
single
server
that
if
it
is,
if
a
server
happens
to
crash
there's
more
factors
that
I
can
depend
on
you're
gonna
be
spending
more
time
wondering
okay,
which
of
these,
which
of
these
parts
is
the
one
that
took
down
my
server
and
then.
Similarly,
if
you
take
it
down
for
a
hardware
change
just
to
just
for
maintenance,
it
will
also
take
down
a
larger
part
of
your
ceiling
power.
A
While
it
is
down
whether
from
a
crash
or
from
a
hardware
change,
if
you
have
80
of
your
power
in
a
single
server,
and
now
you
have
to
go
in
and
change
the
parts
now
you
only
have
20,
also
condensing
that
many
parts
into
a
single
server
also
condenses
the
amount
of
electricity
in
use
by
that
server
problems.
That
would
be
high
heat.
A
So
after
rackspace,
I
also
want
to
mention
power
for
the
power
I
want
to.
First
give
the
examples
of
our
picnic
servers
and
the
maximum
energy
consumption
used
by
them.
I
split
them
up
first
into
the
workers,
the
pc
one
server,
pc2
server
and
c2
server
and
their
watt
usage
at
the
bottom,
their
parts
and
then
also
the
assumptions
I
made
for
their
power
usage.
A
Now.
The
important
thing
to
know
about
workers
is
that,
while
you're
not
actively
sealing
either
cc
or
deals
that
these
workers
can
either
be
idle
or
even
turned
off,
if
you
are
not
planning
on
sealing
for
a
long
amount
of
time
at
that
point,
their
power
consumption
is
not
much
of
a
concern
for
you.
A
However,
these
three
nodes,
the
minor
node
damon
noted
market
node,
will
almost
almost
always
be
running.
The
miner
node
and
daemon
node
unless
perform
getting
maintenance
performed
on
them
should
always
be
running,
because
the
monitor
node
will
have
to
be
doing
its
daily
audit
on
your
sectors,
the
window
post
and
then
the
daemon
node
has
to
keep
communicating
the
blockchain.
Otherwise,
you
won't
be
able
to
do
anything.
A
A
Okay,
so
with
that
in
mind,
this
is
just
some
general
energy
tips
that
we
at
picnic
haven't
used.
First
is
the
very
simple
equation:
amps
times
voltage,
equals
watts,
I'm
sure.
If
you've
ever
taken
a
physics
class,
then
that's
an
equation,
that's
very
familiar
for
you.
So,
even
though
it
is
basic,
it
is
important
because
that's
the
one
you've
been
using
to
calculate
the
amount
of
watts
available
to
you
in
a
circuit,
and
the
last
thing
you
want
to
do
is
trip
over
that
circuit.
A
A
It
also
gives
you
a
important
factor
to
consider
with
rack
location,
because
you
may
find
that
a
single
rack
cannot
power
many
power
hungry
servers.
For
example,
if
you
remember
our
pc2
servers
and
c2
servers
over
both
over
a
thousand
watts
each,
if
you
cram
all
of
them
into
a
single
rack,
it
may
not
be
able
to
support
that
much.
A
A
A
I
won't
be
reading
off
every
single
one,
but
the
two
that
I
haven't
mentioned
would
probably
be
the
network
interface
controller
and
the
storage
jbod
for
a
network
interface
controller
depending
on
bandwidth.
Those
could
be
about
150
700
and
then
the
jbot
enclosure
would
usually
could
be
ranging
from
four
thousand
and
then
you
also
have
the
feel
of
your
hard
drives.