►
From YouTube: CDS G/H (Day 1) - Erasure Coding Update
Description
https://wiki.ceph.com/Planning/CDS/CDS_Giant_and_Hammer_(Jun_2014)
24 June 2014
Ceph Developer Summit G/H
Day 1
Erasure Coding Update
A
B
Sure,
since
I
promptly,
I
can
speak
clearly
as
long
as
nobody
talks
to
me.
That's
weird,
so
a
few
items
regarding
tautology,
the
way
here's
your
coat
profiles
were
supported.
Initially,
where
was
kind
of
bizarre,
because
there
were
no
profiles
initially
etc.
So
there
is
a
planning,
pull
requests
that
adds
support
for
generic
area
code
profiles,
which
would
low
to
define
jobs
for
the
tree-sweets
for
local,
irreparable
codes
or
any
kind
of
variations
of
parameters
that
we
like.
So
this
is
for
the
QA
thing.
B
Then
there
seems
to
be
work
to
be
done
regarding
original
plugin,
optimization
so
far,
I
focused
on
stabilizing
Jerry
you
plug
in
and
doing
some
benchmarking
to
know
how
much
for
put
we
can
have,
but
I
did
not
do
any
work
on
trying
to
figure
out
if
we
could
optimize
something
if
something
was
not
performing
well
enough
compared
to
what
it
could,
because
the
measures,
the
performance
that
we
get
are
good
enough
for
now.
So
this
is
an
area
where
how
we
could
do
something
I,
I,
guess
fairly
easily.
C
I'm
interject
I'll
interject
something
really
quickly
here.
One
of
the
I
think
on
intel
processors
in
general,
JRA
sure,
does
really
well
and
there's
also
the
Intel
stuff
that
much
in
a
minute.
But
the
big
question
for
me
is
what
happens
when
you're
on
ARM
processors,
this
I
think
is
especially
relevant
when
you
look
at
the
ethernet
drives
that
are
sort
of
on
their
way
to
market,
where
you're
going
to
have
these
tiny
little
arm.
Processors
on
the
disk
that
are
acting
as
those
DS
and
you
want
to
use
them
for
cold
storage.
C
B
C
C
B
B
There
also
is
the
question
or
more
information
about
the
maintenance
and
contribution
to
Jerry
sure
and
AF
complete.
So
at
the
moment
we
meant
we
have
a
copy,
the
stable
branches
for
each
of
them
in
the
South
namespace
on
github,
and
they
exactly
match
the
latest
stable
version.
So
we
were
diligent
in
pushing
back
whatever
we
needed
to
for
theft
work
over
there.
B
B
Yep,
that's
it
Oh
is
the
local,
irreparable
code
nah.
I
added
a
link
to
the
benchmark
by
I
just
talked
about
I
did
the
benchmark,
but
I
did
not
analyze
the
call
graph
to
figure
out
exactly
why
there
is
such
a
big
difference
between
4k
and
one
megabyte
encoding
decoding.
There
is
a
significant
difference
there.
It
shouldn't
be
so
loud,
I
guess
doing
something
backward,
but
it's
still
good
enough.
So
I
I
do
not
really
worried
about
it
and
then,
if,
unless
is
here,
maybe
he
could
talk
about
his
prototype,
plug-in
or
aiza.
D
Yes,
do
him
hello,
okay,
yeah
I
said
yes,
I.
Just
today
is
contact
with
the
developer
of
the
ICL
library,
but
in
principle
I
implemented
the
plugin.
This
is
even
a
little
bit
faster
than
the
Giro
Asia.
We
have
a
modern
intel
cpu.
There
was
just
one
question
with
the
decoding
of
coding
chance,
but
he
invincible
answered
me
how
to
do
this
efficiently.
So
there's
not
much
missing
here.
It's
just
not
clear.
D
Also
there
how
you
want
to
take
it
through
this
library
comes
either
aesthetic
won
or
shared
library,
and
the
question
is,
if
you
want
to
compile
this
in
the
repository,
you
just
deal
open
it
as
a
shared
library.
What
I
do
for
the
moment?
It's
more
a
packaging
question,
the
implementation
actually
straightforward.
It's
very
similar
to
the
g
major
one
I.
C
D
D
Yeah
I
mean
this:
is
this
library
some
advantages,
because
in
principle
it
has
no
alignment
requirements,
so
you
can
just
go
to
bite
alignment
but
for
sure
for
the
performs
the
developer
to
a
minute
64
by
the
time
it
is
okay,
there's
also.
The
question
has
to
decide
one
on
storyline
to
64.
Byte
are
a
very
small
objects.
B
Yeah,
well,
it's
part
of
larger
pull
request.
Yes,
it
will
help.
C
B
A
C
C
C
C
C
C
D
I
right
Virginia,
it's
it's
completed
words.
You
can
try
it
if
you
want
so
I've,
okay
made
a
plugin
which
is
competent
with
Firefly
and
I
mean
the
encoding
is
like
ten
and
fifteen
twenty
percent
faster
and
the
decoding.
If
I
get
now
this
pets
ride
will
be
the
same
speed,
so
it
eventually
I've
seen
that
for
this
4k
and
coding
you
do
it's,
it's
really
significantly
better
I,
don't
know
why,
but
I
mean
it's
good
news.
Yeah.
C
That's
awesome,
okay,
so
I
think
that
then
I
thought
the
work.
Then
it's
just
to
figure
out
how
to
embed
the
is
al
piece.
So
I
guess
it's
already
there
yeah,
so
I'm,
looking
at
the
the
source
tree
and
you're
in
your
git
tree
that
that's
assuming
that
the
library
since
is
a
shared
library
installed
on
localhost,
is
how
it
works.
Yes,.
C
Tlie,
that's
the
hard
part,
because
we
don't
want
to
like
make
a
separate
set
of
destroy
packages.
It's
much
simpler
to
statically,
linked
it
in
and
I'll,
say
that's
sort
of
what
Intel
or
might
I
understand
what
they
generally
expect
to?
They
don't
really
I
mean,
did
their
normal
audience
for
these
things
is
you
know
you
can
see
or
whatever
who's,
making
our
next
appliance
and
they're
going
to
pick
this
specific
one
that
runs
on
the
model
of
processor
that,
yes,.
D
E
A
B
E
Well,
yes,
so
well,
first
I
start
working
with
the
key
of
the
application
because
it
was
a
bit
confusing
now
I'm
trying
to
implement
the
rather
see
racial
code,
based
on
the
assumptions
that
a
current
model
class
or
replication
of
I
found
a
bit
confusing
how
to
calculate
the
time
for
repairing
I
exchanged
some
emails
today
with
kyle
and
yell.
I
also
found
this
tool
for
bench
market
benchmark
any
racial
code
that
may
be
it.
It
may
help
to
estimate
the
time
of
repairing,
but
I
didn't
have
time
to
check
that.
E
The
other,
the
other
scene,
so
does
the
assumptions
of
the
current
model
I
found
I
found
them
very
basic
and
not
realistic,
but
but
I
still
I
guess.
It
is
better
to
continue
with
this
kind
of
stuff
assumptions
and
get
some
some
work,
some
work
and
then
try
to
move
to
two
other
more
realistic
models
but
yeah
there.
A
E
B
C
C
B
B
Very
research,
but
display
with
the
pull
request,
is
about
comparing
and
displaying,
but
the
tool
to
do
the
benchmark
itself
is
already
in
the
intercessory.
It's
only
it.
Okay,
it's
not
very
human
readable,
but
it's
machine
consumable
got.