►
From YouTube: Ceph Performance Meeting 2023-03-30
Description
Join us weekly for the Ceph Performance meeting: https://ceph.io/en/community/meetups
Ceph website: https://ceph.io
Ceph blog: https://ceph.io/en/news/blog/
Contribute to Ceph: https://ceph.io/en/developers/contribute
What is Ceph: https://ceph.io/en/discover/
A
Hello
folks
welcome
to
performance
meeting
for
March
30th
2023
Marcus
next
couple
weeks,
so
I'll
be
running
these
we're
still
waiting
for
folks
to
trickle
in
here,
but
we
can
get
started
by
looking
at
some
of
the
parkway
status
and
any
discussion
topics
please
feel
free
to
raise
them.
A
There's
one
new
interesting
new
PR
that
this
week,
which
was
adding
a
zero
API
to
RW
Casey,
do
you
want
to
talk
about
that?
A
little
bit.
B
Yeah
sure
we
talked
about
it
a
couple
weeks
ago
here,
but
it's
the
idea
is
just
to
add
a
a
rest
API
that
doesn't
have
any
back-end
store
so
that
we
can
run
benchmarks
essentially
that
isolate
the
HTTP
front.
End
part
itself
interesting
for
optimizing.
The
the
Boost
front
end,
but
also
for
comparing
work
against
the
HTTP
3
front.
End.
B
I,
don't
think
so.
No
Mark
kogan
did
some
kind
of
very
initial.
Hs
bench
runs
just
to
compare
this
API
against
the
Red
House
backing
with
nvme
kind
of
the
fastest
that
he
could
get
from
that,
and
that
shows
a
pretty
good
Improvement.
B
But
yeah
and
the
in
the
Beast
front,
end
I,
think
there's
some
kind
of
known
overhead,
especially
from
SSL
connections,
so
being
able
to
dig
into
that
specifically
I
think
will
be
good.
A
A
Okay,
in
terms
of
other
updated
PRS,
we
have
more
progress
on
the
booster
accumulation
idea
that
Mark
has
proposed.
This
is
because,
like
you
always
reviewing
it,
then
Mark
was
trying
to
update
it
a
bit
more
Mark's,
not
here
this
next
couple
weeks,
so
we'll
probably
discuss
it
more
when
he's
back,
but
the
idea
is
essentially
to
try
to
batch
up
more
key
Value
Store
updates.
Thank
you.
Any
committed
them
for
potentially
longer
intervals,
based
on
different
configurations.
B
A
Mm-Hmm,
the
other
one
next
one
is
I,
know
we're
properly
overwriting
the
marksdb
writable
file,
allocate
method
which
is
addresses
a
fragmentation
issue
with
blue
FS,
like
this
is
all
approved
now
and
just
needs
some
QA.
A
Now
we
have
a
new
Crimson
performance,
Suite
I
think
this
might
have
been
discussed
in
the
past
weeks,
but
it's
not
looking
looking
quite
good
passing
consistently
and
we
can
start
collecting
data
and
seeing
how
Amazon
is
doing
over
time.
A
And
finally,
we
have
the
very
large
PR
around
updating,
M
clock
recovery
and
backflow
costs.
There
is
some
you've
seen
where
back
film
recovery
was
going
much
slower
than
expected.
A
This
ended
up
being,
especially
with
small
object
workloads.
This
ended
up
being
due
to
the
the
way
that
mclach
was
treating
the
cost
of
recovering
it,
because
I
given
object.
A
So
there's
a
few
changes
here
around
tuning
the
our
cost
for
a
given
PG,
based
on
the
average
size
of
objects
in
that
PG,
instead
of
a
PG,
is
mostly
used
for,
like
RVD
objects.
There
are
all
four
megabytes:
it's
going
to
be
pretty
much
the
same,
but
for
rjws
ffs,
the
object
size
can
vary
quite
a
bit
between
different
pgs
and
different
pools,
so
yeah
or
when
we're
scheduling
the
recovery,
taking
that
into
account,
as
we
don't
know
exactly
which
obviously
we're
covering
at
that
time.
A
At
that
point
of
scheduling,
we
have
to
kind
of
rely
on
the
average
there
and
then
for
that
I'm
updating
the
recovery
costs
for
different
kinds
of
operations,
so
that,
if
the
within
recovery
we're
only
accounting
for
things
once
not
multiple
times,
that's
that's
what
the
pull
up
versus
push
out
for
flight
changes
are
for
I.
Believe
there's
some
other
changes
here
to
make
the
configuration
overall
a
bit
simpler
and
more
more
sensible.
A
But
this
is
all
still
under
review,
but
I
could
write
it
right.
I
can
say
we're
taking
a
look
at
the
review
and
some
more
updates
from
sweetheart
there.
So
hopefully
this
will.
You
can
finalize
the
next
few
weeks.
A
That's
it
for
the
PR's
General
announcement,
I
guess
for
switching
this
meeting
to
jitsi
and
Elsa
meetings,
Jujitsu
General
starting
next
week
since
blue
jeans,
and
so
the
okay
look
at
the
calendar,
invite
for
the
updated
links
for
this.
One
should
be
pretty
simple
to
join.
We've
been
using
it
for
the
self
leadership
team
meetings
for
eign.
A
That
was
all
we
had
on
the
Nintendo.
So
far
are
there
any
topics
for
discussion
folks
would
like
to
bring
up.
A
Any
interesting
papers
or
things
that
people
have
read
or
people
might
want
to
look
at.
A
C
I'm
sorry
I
think
I
was
not
even
being
recognizing
sorry
I.
Think
I
I
had
a
meeting
due
like
safe
research,
weekly,
but
I
was
not
able
to
yeah
I've,
been
work
on
it
and
and
like
think
upon
those
things.
So
I
thought,
like
maybe
we'll
come
now
and
talk
about
it.
How's
things
going
on
and
stuff
in
general
as
well
familiar
with
them
as
well,
so
I
think
Mark
added
some
ideas
to
the
nuclear
part.
Tech
performance.
C
I
think
also
there's
one
more
meeting,
the
that's
being
scheduled
now
for
Blue
Store
itself,
I'm,
not
sure
about
things.
Are
there
so,
like
he's
a
is
this
meeting
being
used
to
discuss
all
the
PRS
and
you
know
for
booster,
we
have
that
meeting
separately.
A
Yeah
yeah
I
guess
I
haven't
been
going
to
the
blue
star
meeting
since
it's
a
bit
early
for
me,
but
essentially
it's
more
around
the
looser
specific
improvements
and
design
from.
C
Yeah,
probably
I'll
share
in
a
mail
with
you,
but
their
performance.
Basically
that.
C
C
Maybe
yeah
and
staying
up
with
you
on
the
mail
regarding
that
and
yeah
get
going
on
that
as
well.