►
Description
https://wiki.ceph.com/Planning/CDS/CDS_Giant_and_Hammer_(Jun_2014)
25 June 2014
Ceph Developer Summit G/H
Day 2
Systemtap/LTTng static markers (update)
A
Okay,
so
at
the
moment
there
are
kind
of
two
paths
free-floating
around
there's
a
demand.
A
A
With
mine
I,
take
your
system
tap
mostly
Iran
them
I,
don't
actually
think
the
choice
of
tracing
system
is
particularly
important.
I
think
we
probably
support
both
it's
just
a
way
of
carrying
information
from
stuff
out,
but
I
was
marker,
stood
in
bracing
the
resource
contention
and
tying
it
out
back
to
and
off.
So,
ideally,
once
that's
done,
every
mutex
in
the
OSD,
when
you
take
it
you're
obliged
provide
an
OP
that
we
can
associate
with
during
the
duration
of
the
longer.
B
A
B
Okay,
yeah
I
mean
it
I
told
you
this
before,
but
my
only
concern
is
that
it
would
be
nice
if,
whatever
trace
points
annotations,
we
put
in
the
code
are
general
yes
generalized
to
most
systems.
I
mean
it
does
matter
that
much
now,
cuz
aren't
that
many
of
them,
but
before
we
go
sprinkling
them
all
over
the
place
with.
A
Proposed
DSM,
it
basically
replaces
mutex
with
track
mutex
and
the
lock
and
unlock
methods
you're
obliged
to
provide
a
track
out,
and
there
are
here's
some.
There
is
an
annotation
method
associated
with
each
opposed
with
the
thing
that
outputs
to
system
tab.
So
then
you
just
add
another
one
to
that
one
place
that
also
want
to
put
it
still
TT
in
G,
so
for
an
OSD
mpg
you're.
Not
if
you're,
not
a
meeting
system
tap
event
because
obliged
to
provide
off
when
you
take
a
lock,
basically.
A
I
mean
you
provide
trace
points
that
are
a
trace
point
name,
followed
by
a
bunch
of
entities,
information
which
gets
its
lipid
by
the
other
side
of
the
call
give
me
whatever
receives
LT
Indians,
both
PPMD
fixed
they'd,
really
get
this
list.
It
could
be
a
way
of
shuffling
bubbles
for
twenty
two
point:
oh.
A
It's
a
little
more
than
that,
because
they
they
have
the
ability
to
do
hot
passing
of
the
code.
So
if
you,
if
you
use
something
like
say,
I
want
to
trace
only
those
things
that
are
the
apps
that
that
branch
that
says
I
want
you
to
not
do
anything
as
the
app
happens
and
stuff
not
in
with
grace
camp,
doesn't
actually
how
patches
left
side
with
a
with
with
compiled
native
code
that
is
capable
of
making
that
actually
quite
snappy.
B
Yeah,
so
I
am
Adams
having
he's
working
on
getting
blue
jeans
up
and
running,
but
I
think
so
no
is
looking
at
understanding
where
the
time
is
spent
in
the
OSD
IO
path.
Wyatt
opts
for
taking
as
long
as
they
are
Adam
I
think
is
looking
at
doing
tracing
on
the
client
side
and
I.
Think.
The
idea
is
that
you
would
have
a
you
know:
a
standard
build
of
the
client
library,
whether
it's
labret
dose
or
the
VAR
BD
and
be
able
to
you
know,
subscribe
to
a
stream
of
trace
events.
A
A
Little
the
reason
for
all
the
complexity
and
system,
tab
and
ulti
tedious.
You
want
to
be
able
to
net
to
dynamically
enable
these
branches
at
huge
numbers
of
prisons
in
the
code
there
being
any
overhead
at
any
other
time.
It's
just
getting
the
things
that
we're
gonna
send
over
the
wire
as
a
message
anyway.
I'm
not
sure
that
baby.
B
B
A
B
Yeah
in
that
case,
if
you
wanted
to
make
it
our
beta
straight
level
trace,
but
you
would
want
to
be
able
to
do
an
RB
D
level
trace.
So
you
can
see
the
operations
above
the
client-side
cache
and
also
like
within
the
independent
of
the
striping
strategy
that
our
VD
is
doing
or
the
cloning,
and
all
that
stuff
I
mean
I.
Think
one
that
this
again
would
sort
of
a
little
bit
about.
But
Adam
was
focusing
on.
B
One
of
the
general
challenges
is
that
you
have
all
these
people
deploying
clouds
and
running
virtual
machines,
and
they
want
to
know
what
we
want
to
know.
What
kinds
of
workloads
are
actually
seeing
and
figure
out
what
generalities
exist
and
what
don't?
Because
it's
you
know
who
knows
what
people
are
running
in
the
clouds
and
what
beyond
saying
that
there
a
lot
of
small
is
from
those
most
virtual
machines.
It's
hard
to
say
know
exactly
what
that
workload
profile
looks
like.
B
B
B
We
talked
about
that
at
fast
a
couple
years
ago,
that
was
I
think
was
actually
a
group
at
Harvey
Mudd
that
worked
on
that
they
took
well.
They
took
block
level
traces,
but
the
output
of
that
was
a
file
bench
config
file
that
modeled
the
same
the
same
characteristics
sent
to
the
northen,
yeah,
yeah
and
I.
Think
that
that's
actually,
sir,
the
direction
that
that
we
eventually
like
to
go
in
right.
So
we
can
actually
just
describe
the
nature
of
the
workloads.
B
B
Well,
for
the
next
one
we
do
on
how
am
I,
though,
because
it's
a
copy
on
the
shared
flag
right,
mm-hm,
so
I
see
min
is
here.
We
could
jump
to
the
copy
and
read
clones.
You
wanted
to
put
the
whole
schedule
right
through
a
blender
thing
to
do.
Alright.
Is
there
anything
left
to
cover
off
on
the
system,
tab
stuff
or
should
I
shut,
that
down.