WWDC2004 Session 602

Transcript

Kind: captions
Language: en
morning before we start this session I
want to talk a bit about the session
that's gonna be coming after this
because Exxon is a brand-new technology
and we have tons of things we want to
talk about we split the content into two
sessions the first one is this one
introducing X N and the second one is
the one right after this 1030 in this
room X and in depth so if you're here to
learn about what Exxon offers and what
it takes to deploy and manage it what to
learn how to program for it we highly
recommend that you attend both sessions
and you just wanted to quickly review
what we're going to talk about in these
two sessions so in this session we're
going to start off by talking about why
you may need a Stein file system for are
the benefits and we got to spend a
pretty good chunk of time talking about
all the key features of Excel everything
from the basic file system to the
management to performance the
cross-platform support and we're gonna
cover what it takes to deploy exam
basically going to go over the basic
hardware configuration and with a few
examples and then to conclude the
session we're going to talk about
different software and hardware product
apple offers to do the same and also
build different solutions on top of that
so that should keep us busy for next
hour to hour and a half and then at
10:30 we're going to talk about how
Excel works from engineering perspective
so for example we're going to talk about
things like what gets passed on wire
when you make a filesystem call and
we're going to spend a lot of time
talking about the administration of X
and in fact we're gonna have a demo of X
and admin software using it to set up
Sam from scratch or almost from scratch
so that should be pretty fun right and
we're going to talk about different
aspects of administration such as
command-line tools then we're going to
talk about a few tips for developers
thinks they should watch out for when's
accessing some volumes example in sync
such as scalability and locking and
other issues and the last week we're
going to talk about excellent developer
API we will have a demo for a couple
sample application using that API
and it's pretty visual so it should be
pretty entertaining so is that I would
like to introduce the Speaker of the
first session Tom Morgan director of
server and storage software from
worldwide product marketing good morning
everyone so its first thing in the
morning and I just I was playing hockey
last night until late so I'm going to
try to move along quickly and hopefully
keep myself going and keep all of you
going as well I have a few questions to
ask right from the beginning how many of
you are familiar with storage area
networks in general okay that's good
how many of you have actually deployed a
storage area network all right how many
of you put a SAN or a cluster filesystem
on that storage area network and how
many of you have written a cluster
filesystem yeah
see I knew Ted Klein was in the audience
so I knew I could find someone who had
done that at least so I'm gonna take you
through kind of an overview of storage
storage area networks and anxn give you
some background information on xn and
storage area networks in general just to
make sure that we're all using the same
terminology because in the industry the
terminology somewhat varies from vendor
to vendor so take you through what we're
doing I'm gonna give you some
information on our cluster or our sand
file system and then I'm going to
conclude by taking you through the range
of products we now offer for IT
deployments so let's let's jump right in
real quickly this is xn and we refer to
it as the sand file system for Mac OS 10
and I can't say this will be the final
mock-up of the box but this is this is
our product we introduced this at NAB in
April it actually won an award as best
of show at NAB people were very excited
about it we demonstrated it at NAB I
think we had 6 terabytes of storage in a
single volume shared among 5
six different systems these were desktop
systems and we were doing video
real-time online video editing from
those systems so it provides some really
really great performances some great
features okay
there we go so as a background err for
some of you I know some of you this will
be very trivial on the other hand for
some of you I just want to bring you all
up to speed on sort of why we decided we
needed to have a sand file system or a
cluster file system and I'm gonna take
you through the basics of storage
architectures but I'll move along
quickly so we don't don't more some of
you in the audience so direct attached
storage or DES this is very very simple
everybody does this it's plugging an
external storage device into a server or
a desktop the desktop or server mounts
that storage device and of course if
it's a file server it can actually share
that storage to clients across your
network and it does that with typical
network file sharing protocols so it's
pretty straightforward you can use fibre
channel or a scuzzy interconnect between
the server of the desktop of course you
could use firewire if you wanted to as
well there's a number of direct attached
storage interconnects today you could
use USB and share that if you wanted to
now the one thing about this is that the
file server becomes a potential
bottleneck for the performance of that
overall system the file server is the
only way to access that external storage
and the file server is the final arbiter
on who gets access to that storage so
everything goes through that one file
server now not only is it a performance
bottleneck but it is a potential single
point of failure as well now what came
out and came into vogue a few years ago
when as devices network-attached storage
the reason these are very interesting is
that these are purpose-built appliances
for sharing storage over the network so
in a sense it's just like a direct
attached storage device except that it's
storage with an embedded file server
it's a good way to think about it so it
has some really great advantage
one of which is it's a queue typically
Nazz devices are extremely easy to
deploy very quick to get them up and
running and that's because they are
single purpose devices their whole point
is to share storage on the network at
the same time they do suffer from some
of the same performance and scalability
limitations as direct-attached storage
now there's there's a whole range of
very sophisticated things that some
companies offer in terms of storage
virtualization services where they can
actually make multiple NAS devices
appear like and and on the network
appear like a single device or they can
eliminate some of these single points of
failure but it is fairly complex to
deploy the storage virtualization
software to accomplish this now in a
storage area network this is where
multiple storage devices are attached to
a fiber channel network that's really
what a storage area network is all about
now fibre channel today as well as ice
Guzzi you could use a different type of
network called ice cozy this is scuzzy
protocol over IP over Ethernet but for
all intents and purposes today when I
talk about a San I'm going to be talking
about a sand-based on fibre channel so
we have multiple des storage devices
connected to a fiber channel switch and
then we have multiple servers that are
able to access those storage devices now
a key thing is that they essentially
access the storage devices and in a
sense they share the storage devices but
they don't share the data on the storage
devices I'll get into a little bit of
detail on that right now
so why a SAN filesystem well as I said a
SAN you have multiple computers
accessing the same storage device an
example here we've got an X Server aid
connected to a fiber channel switch the
Xserve raid presents to raid controllers
hence the two fiber channel lines coming
down from the top and the fiber channel
switched to the 3x herbs that have shown
here these three excerpts are connected
via fibre channel to the fiber channel
switch and they can see the Xserve raid
each one of them can in fact they could
take that storage from the X
raid and they can reshare it using
network protocols such as AFP SMB or NFS
across the network to these clients so
they can do the exact same thing that an
AZ device or direct-attached storage can
do however the challenge with this
before you have a SAN file system is
that each raid set is effectively a
volume and the volumes are connected
separately to each one of the servers so
you have a B and C up and they're
denoted by the colors and a B and C are
mounted by each of those servers only
one server can mount one filesystem so
the server could mount multiple file
systems but the file system or the a or
the B of the C can only be mounted by
one server can't be mounted by two
servers at the same time so each raid
set is mapped to a server that's a
process called provisioning in the sand
world and that's what that's all about
the reason a SAN file system is
interesting is you eliminate these data
silos keep in mind that server a over
here only sees the data in volume a
server B only sees the data in volume B
server C only sees the data and volume C
that's before a SAN file system after a
SAN file system you can actually have
data a B and C as denoted up a hot Bob
in a single volume so this would be a
single three three and a half terabyte
volume on the Xserve raid and that
volume can be seen by each of these
servers so this is kind of a subtle
difference between network-attached
storage and sans in that the filesystem
used for a sand can actually the Mel
volumes created by that filesystem can
actually be mounted directly with direct
block level access by each of the
servers or desktops that that's attached
to the sand so multiple computers all
three of these computers can read and
write to the same volume at the same
time and
is block-level access very very
different from in a sense from when you
have a network file system it may be
sharing a file system to multiple
computers but final arbitration
block-level access to the volume that
it's sharing is arbitrated by the file
server itself by that single file server
this is very different and of course it
could we share that volume to all of
these clients as well so now we have
three servers able to share the data
that's on that volume out to the network
now the nice thing about a SAN file
system is it's very very easy to scale
so you want to add more storage you've
gone past three and a half terabytes and
you want to go to five or six terabytes
of storage you literally add the storage
in you can grow the file system and all
of the systems that are attached
directly to the file system simply see a
larger volume with more space for more
data now again the other thing you can
do is you can add more servers to the to
the san as well and share the data in
the volume through more file servers so
you can literally get more Network
performance out of the sand than you
would get with having a single server
sharing a volume because now you have
three four or five or six as we show
here servers able to share the data
that's on the same volume so that's why
a sand file system is very interesting
so a little overview of X and X sands
for high-performance storage networking
so it's very very fast in fact when you
setup an X and system properly
configured it'll be I don't say this is
a lot of you it'll be much much faster
than the internal disk drive on your
system in fact you will get whatever the
storage architecture that you deploy can
give you X n is very very lightweight in
terms of its over overhead on the
storage infrastructure
it has a high availability architecture
so you can imagine a system like this is
going to deploy be deployed in
mission-critical environments we've done
a lot of things to make sure that when
you deploy it that way if any single
failure happens somewhere on the system
we've got you covered and I'm going to
take you through some of that as well
xn provides some really really
interesting volume management tools and
that's where we get into some of the
nomenclature I mentioned and I'm gonna
walk you through a little bit of the
volume management tools and show you
what we're talking about there and then
later on I know the guys are going to
show you how to actually set up a sand
volume accent is going to be easy to use
and when I say easy I don't want to be I
don't want to be you know really
lightweight about this whole thing
obviously you need to know about storage
area networks to get one built and one
running even before you put X and on it
you're going to need to know a little
bit about sand file systems and how to
deploy them but for most of the
mainstream configurations where we're
hoping that our management tools will
help you get those up and running in a
nice optimized way without having to
know - too much detail about the innards
of the sand file system so acting a
little bit about that and and we have a
great story when it comes to
compatibility and interoperability and
I'm gonna run you through all of that
because no sands solution stands on its
own no sand solution from a single
vendor really stands on its own unique
partners involved to be successful in
this space so what is it it's a 64-bit
cluster file system so it's a full
64-bit cluster file system now on on
Panther that means that you'll get up to
16 terabytes values that's the
limitation built into Panther today on
Tiger that limitation goes away with X
and so on Tiger you'll be able to be
able to build very very large
if you don't think 16 terabytes is large
enough
very large volumes now to give you some
idea of what's happening in the video
space there are customers that have
asked us for over a hundred terabytes of
storage in a single volume and so we're
gonna take a shot at meeting their needs
as we go through the deployments of xn
you can obviously have billions of files
up to 64 concurrent clients can access
an xn volume so what does that mean that
means that you can have up to 64 systems
I showed you six systems and the largest
deployment I showed here you can have up
to 64 servers or desktops or mix of the
two accessing the same volume direct
block level access over fibre channel to
that one volume and of course you could
create multiple volumes as well and have
a different 64 at you know connect to a
different set of volumes that does not
mean that you are in any way limited by
the number of network clients that are
coming in over AFP SMB or NFS and
connecting into a file server that is
directly connected to the sand we're
only talking about here the number of
machines that are directly connected to
the sand and of course it gives you file
level locking and what this really means
is that on a single volume you can have
multiple readers and multiple writers so
you can have multiple computers writing
to the same volume at the same time and
you can have multiple computers reading
from the same volume and in fact in this
case reading from the same vial at the
same time the only limitation if you
call it that is that you can only have
one computer writing to a particular
file at a given time so from an H a
perspective do a couple of things
there's a thing called a metadata
controller and I'm going to show you a
little bit more about what that's all
about in subsequent slides but the
metadata controller is like the traffic
cop for the sand what the metadata
controller does is that anytime a
computer on the sand
wants to open a file read a file write
to a file it first talks to the metadata
controller and it says you know can I
open that file I'd like a read/write
access to the file XYZ and the metadata
controller checks and says ok no one
else has got read/write access right now
so here you go you can have readwrite
access and you know here's where the
file is located but note the difference
from a network file system it doesn't
say here's the file
it says here's where the file is located
and then the computer that requested the
file goes in over fibre channel directly
to the storage at the block level and
reads or writes to that file so that's
what metadata controller failover is all
about is providing a secondary metadata
controller you can imagine if the
metadata controller fails then you've
got a real problem on your network on
your sand because nobody knows if
they're allowed to have access to a file
where the file is located and so forth
xn comes out of the box with support for
metadata controller failover if any
metadata controller on your sand fails
the xn computers the computers running
XM will get together and they will
nominate another one of these computers
to take over the duties of the metadata
controller they will do that
automatically and that will happen
within seconds of a detected failure and
the metadata controller so that's what
metadata controller failover is all
about one other added note about it
because we ship xn with the metadata
controller software built in unlike a
lot of other sand vendors any X and
system connected to the sand can be the
metadata controller so we support
something called cascading failover this
means if the second metadata controller
if that person who stumbled into the
machine and kicked the power cord out
stumbles into the next metadata
controller and kicks the power cord out
then the next system on the sand will
take over and so as long as any one
system any one xn system is up and
running you have access to your data
we think that's extremely important with
these systems of course the file systems
journaled you'd expect that and that's
why we can actually recover the file
system within seconds of a failure the
other metadata coentrão takes over
accesses the file system runs the
journals and you're up and running
within seconds we support something
called fibre channel multipathing so
some of you may have purchased an X
serve and you may have gotten the fibre
channel card that goes along with it we
have a fiber channel host bus adapter or
HBA
that goes into the machine so that you
can connect a fiber channel cable into
the excerpt or into a Power Mac g5 the
thing is that that card actually
supports two connectors and you could
actually run both those connectors to
the fiber channel switch where they can
see the storage and the reason we do
that is that we use both of those
connectors in X and this is what
multipathing is called it refers to and
we can use both those connectors to send
requests for packets or blocks of
storage and receive those blocks of
storage so you can actually get a
theoretical 4 gigabits per second each
fiber channel connection provides 2
gigabits per second you can get a
theoretical 4 gigabits per second into
an X served g5 for example they keep in
mind that that's theoretical because
you'd have to have a storage
architecture on the other side of the
sand you'd have to have hard drives and
raid controllers and so forth that can
actually deliver 4 gigabits per second
but if you had that you could get it in
there the other thing with multipathing
is if the one of the past breaks so
someone steps that same guy who
stumbling around your data center
actually breaks one of the fiber cables
that are running to your storage for
joining to your switch the other
connection continues to work and the
filesystem continues to function and
that computer continues to have access
to the data that's on the on the storage
so that's fiber channel multipathing
there's one last feature now this
feature is interesting because it
actually requires that
you those of you are a developers be
involved in this and this is called
bandwidth reservation now I want to kind
of describe in a sense I mentioned that
you know if you have theoretical 4
gigabits per second you can get that
into the computer if you have the
storage architecture to support it it
can get very expensive to say support
for gigabits per second into every
computer that you have on the sand those
of you have tried to do some of this
understand what I mean you can buy a lot
you have to buy a lot of storage to be
able to do that you have to have a lot
of disk drive spinning in order to be
able to deliver data like that so when
most people design their sand they
design it to deliver the aband with that
they need for typical operations they
look at the mean they look at what
they're usually doing every day day to
day so if you're a video person and
you're doing SD video you'll design your
sand to give you multiple streams of SD
video and audio and it's some type of
audio to to each of the systems that are
connected each of the editing base for
example that are connected to your sand
in the same sense if you're setting it
up for file storage you'll take a look
at each of the servers that you have
sharing the files across the network and
say how much bandwidth storage bandwidth
do I need to provide for each of these
machines you'll sum it all up and you'll
try to create an environment that
delivers that from a storage
architecture perspective now in some
cases someone may walk in and say you
know what and this this happens in the
video space a lot you walk in and
suddenly someone wants to capture HD
video right into the sand now that's a
critical operation it's going to take an
hour or two or three depends on what
you're doing and you want to let them do
that
and you want to make sure that while
they're doing that if some editor comes
in to another station and sits down and
decides to edit some stuff that they
can't mess up this critical capture off
to rate operation that's happening in
one of the stations so the way that
works is that the software managing the
capture operation
can be written to use our bandwidth
reservation api's we have a set api's
for this so it can be written to use
that and so the administrator when they
set up the capture operation can say I
want this particular operation in this
particular file handle for example to
get X Y percent of the available storage
bandwidth so you know what your storage
bandwidth generally is and you say well
I want to make sure that this guy always
has 65 percent of the available storage
bandwidth no matter what these other
stations try to do fail them if you have
to make sure this guy always gets
everything he needs and that's what
bandwidth reservation allows you to do
you can allocate a percent of the
available bandwidth to a specific file
handle you do that for certain critical
operations that don't happen very often
so that you can literally get more
bandwidth out of your system in a sense
it's not going to give you any more than
you have but you can deliver a certain
amount of that bandwidth directly to one
operation when you need to do that sort
of thing so that's what bandwidth
reservation is all about from a volume
management perspective we support a
range of things and I'm gonna walk you
through a little bit with this with some
pictures storage pooling so we group
multiple storage devices together really
multiple raid devices typically and
we'll you'll hear me with others refer
to them as Lunz logical unit numbers
alone is simply a logical grouping of
storage that's hidden behind some form
of RAID controller typically and and so
for example with the Xserve raid as many
of you know it has two raid controllers
in it seven drives on either toss ID
attached to each of the raid controllers
if you took the seven over here and just
did they raid five across the mall to
that RAID controller the seven over here
raid five across them to that controller
when you looked at it across the fibre
channel network with X and soar with X
and admin tools you would see two Lunz
and these two Lunz
grouped together and something we call
the storage pool in fact you can group
multiple Lunz
across multiple devices into multiple
storage pools in fact so I'll show you
how we can do that sort of thing and
what that's all about we can dynamically
scale volumes as I mentioned earlier so
you can grow these volumes we support
things like volume mapping and
affinities I'm gonna go into a little
bit of detail on this
you can decide with the xn admin tools
which systems connected to the sand are
able to mount which volumes and you can
have that mount occur you can force that
mount or force that amount rate from
your admin console and as I mentioned
the admin console works remotely so you
can actually run that on a power book or
an iBook and connect into your sand
remotely and take care of that type of
management affinities I'll go through
it's better to do a diagram for that
because it becomes really really clear
what affinities are all about with the
diagram and we do have offline
defragmentation tools now the file
system is designed and by design
designed to minimize if not eliminate
fragmentation but there are of course
certain operations that can happen on a
file system that are sort of out of the
out of the fringes that can cause some
amount of fragmentation and so we
include the defragmentation tools along
with a product so that you can run those
if necessary on the file system so let's
talk about how we create a volume I
mentioned the Xserve raid I build a raid
set as I mentioned from say those seven
drives I format it or initialize it I
should say raid five and I do that with
the raid admin tools this is we have to
be very clear about what we tools we use
with xn because you will not be using
Disk Utility with xn ok Disk Utility is
for creating hfs+ or ufs volumes and
doing software raid X and has its own
volume format and so when you create an
X and volume you'll be you
using the ex an admin tools to do this
so we create a raid 5 LUN we haven't got
a volume yet but we've got a LUN it
shows up on fibre channel we can take
multiple Lunz as I mentioned and put
them into a storage pool and the storage
pool allows you to group multiple LUNs
that are similar to each other so take
multiple raid 5 ones each one say is one
and a quarter terabytes in size and you
group them together
what xn does is automatically stripes
data across the storage pool so now you
have this logical grouping of storage
this logical pool of raid 5 loans you
can actually create multiple pools and
put them into a single volume and on the
example I'm showing here we have a raid
5 storage pool made up of one or more
raid 5 Lunz and a raid 1 or a mirrored
set a storage pool made up of one or
more raid 1 or mirrored sets ok and
we've pooled the two things together we
haven't pulled them together we've
grouped them together in a single volume
now some of you out there saying well
that's kind of cool but where does my
day to go right I mean does it go on the
raid 5 hardware does it go on the raid 1
hardware just exactly where does it go
and that's what's kind of interesting
about affinities so we all know the raid
5 in theory at least would be a little
bit faster than the raid 1 in typical
hardware storage hardware the read one
however is mirrored so it's actually the
most protected type of storage that you
have it's mirror it's a mirrored set so
you have fastest and most protected now
what you can do is you can actually
through affinities associate those
storage pools with directories or
folders in the finder or folders in that
volume so if you had a couple of
desktops hooked up as I show in here you
can literally provide the user
with two folders in this case I've
called the mo fastest and most protected
you could call them work in progress and
final work for example and the work in
progress could go to the raid 5 and the
final work could go to the raid 1 so in
a sense in reality affinities will
really help you set up interesting
workflows across the sand and move data
around the sand into the type or
category or class of storage that is
relevant or most useful most appropriate
for the type of data that you want to
store in it so this is really
interesting feature of X an and and I
invite you to check that out as well
from Road administration perspective as
I mentioned fairly easy to configure we
do real time monitoring will do event
notifications will do things like email
you or page you in the event of a
metadata controller failover a fiber
channel problem a quota problem we
support usage quotas we have both soft
and hard quotas and and we do have
integration into LDAP and this is where
I'll take a little a little digression
here if you'll allow me this so from
from a sand perspective the filesystem
sits underneath the standard sort of API
set filesystem API set in Mac OS 10 and
Mac OS 10 server so for all intents and
purposes it just looks like another
locally attached internal if you will
volume on your system or drive on your
system and this is really really
interesting for a couple of reasons now
with a couple of caveats to that
internal drive looks and behaves mostly
like ufs follows POSIX lock locking
semantics and is case sensitive so all
accent systems
we'll be case sensitive and we are
relying on all of you to do the proper
thing from a locking standpoint in your
applications so when you open a file for
a user for readwrite access open it and
typically in carbon this takes care of
itself automatically but if you're doing
a cocoa app or a UNIX application open
it and assert the POSIX lock for that
operation and we will adhere to that and
use that in xn to make sure that your
data doesn't get trampled by some other
user now so that's where xn sets and the
reason that's interesting is that on our
systems that means that you can control
access to Xen to the folders and the
files in X and using a directory server
so when the person logs into the machine
or when they remotely attach to a server
and go in and using say a file service
they have an identity a uie D that has
been given to them by the directory
server at login time and all of the
machines connected to the directory
server in other words all the machines
on the sand connected to the directory
server recognize that UID and we'll use
that UID to determine whether or not
that person has access to a folder or
file on the sand and because xn is
coming out in the fall jump ahead a
little bit that xn is coming out in the
fall right now we're adhering to POSIX
permissions but not Ackles
on X and so we use the permissions model
that's built into Panther for that but
that gives you some control over who has
access over data in X and gives you full
integration with an LDAP directory
server and for that matter
Kerberos infrastructure as well because
we're sitting as I said below the
filesystem layer in the in the operating
system
so I mentioned the complete sand
solution and mentioned that you need
partners to really pull this off so from
a hardware perspective we offer Xserve
raid and we qualified on Xserve raid
they'll be qualified with the Xserve g4
and g5 it will also be qualified with
the Power Mac g5 and with Power Mac g4 s
back to duel 800 or better we have our
own fibre channel PCI card that we're
qualifying the system with as well it's
going to be qualified with our
professional software including Mac OS
10 Mac OS 10 server and many of our
professional applications we have worked
with the fibre channel switch vendors
and we're testing with these switches
there's a list of switches on our
website that we're testing with from
brocade cue logic and Emulex so check
those out and from a data management and
interoperability standpoint we have
worked with Adi C to ensure that you can
add windows linux and other unix systems
to the sand and adi c is supporting Xsan
with their store next file system other
store next that's the compatibility is
with their store next file system and
they're also supporting exam with their
store next management suite now what
that means is that you can use their
data management tools as a foundation
for a whole for your for something
called some of you would be familiar
with this but something called
information lifecycle management and
this is policy based storage management
and they have some really really great
products to do this and you'll be able
to use those with X and to do things
like I mean I think of it as backup on
steroids but basically to be able to
migrate your data from tier 1 to tier 2
storage and migrate it all the way back
to tape so and to automate a lot of that
process
so just a little view of how this can
happen you can literally put Xsan on say
a Power Mac g5 and Xserve g5 have that
connected in it sees the sand volume
there's a metadata controller so X ends
up there running on another excerpt g5
as well but you could literally take a
couple of servers running Windows or
Linux and have those servers connected
into the sand as well now those machines
would be running the store next file
system from Adi C and you can you can
purchase it from them it will install
and run and allow you to see the EXCI
involuntary
to back up and maintain all of the data
in your sand as well so this is a full
complete solution that we're offering
here now take a drink of water here but
I want to take you through quickly some
example deployments and then the kind of
walkthrough of some example deployments
I also want to show you how you
specifically how you set up some of the
infrastructure for xn now in the later
session they're gonna get right under
the covers and show you more of this but
I want to give all of you kind of an
overview of how of all is that all of
this works so when you're setting up an
exon you need a storage device such as
the next serve raid you need a fiber
channel switch you need obviously the xn
systems the systems that are going to
access the storage and this will be
servers or desktops and you run X n on
each of these and they all connect in
over fibre channel to the storage area
network one of these systems you're
going to denote or decide is going to be
your metadata controller remember that
traffic cop for the sand so fiber
channel network you get high-speed
you'll get on each fiber roughly well
roughly you'll get their scope for two
gigabits or 200 megabits per second
and we have numbers up on the website on
what the Xserve rate can actually give
you sustained but it's on the order of
180 190 megabytes per second sustained
throughput so and correct me guys if I'm
wrong on that one but that's my
understanding of those numbers so you
get that out of the Xserve raid and you
can of course provide that to each of
those machines now each of these systems
I want to be very clear about this need
to have a copy of xn running on them so
you buy four copies of xn you put it on
the four systems install that get a
configure using them X an admin tools
select one machine as the metadata
controller decide and wait which of the
other machines should actually be the
backup or standby metadata control or if
the metadata controller fails and you're
on your way
well sort of there's a back-channel
network that needs to be deployed with X
and we refer to it as the metadata
network now I don't want to confuse you
with the metadata that's associated with
the spotlight technology
well metadata we're talking about here
is filesystem metadata file system locks
and locations and all of that sort of
stuff so the filesystem metadata is
moved around on a separate metadata
metadata network now this is just a
tcp/ip Ethernet network that you set up
between all of these systems and what it
provides is really out-of-band
communications and the reason this is
important and the reason that's
architected this way is that what you're
sending back and forth was metadata as
small packets so small requests I need
to open up this file ok you can open up
this file fibre channel is really
optimized for large chunks of data and
so if we were to take all the metadata
operations that would happen on a large
sand and run them across fibre channel
we'd be bogging down your fibre channel
network where all your storage
performance is
with metadata communications so to
optimize around this we put the data the
big chunks of information in on the
storage network on the fibre channel
network and we put the metadata on a
separate network
now the actual metadata actually lives
on the sand but it's accessed and cached
by the metadata controller and then it's
controlled and access to it is
controlled by the metadata controller so
this separate out-of-band network now
I'll have a lot of people have come up
to me and said so do I have to use a
separate network for the metadata
network and the answer is well it's IP
you don't have to use one but if you use
your local area network for example and
someone initiates a large FTP transfer
across that network for whatever reason
that may inhibit some of the metadata
traffic on that network and that may
slow down your file system and almost
guarantee it'll slow down your file
system because that will delay those
packets being sent to the metadata
controller so we recommend very strongly
recommend is that you use a separate
network for this metadata controller
network now the good news is that you
can do that and you can still share the
data over Ethernet when you're using for
example an X or g5 because we do have
two Gigabit Ethernet connectors on the
back of the Xserve g5 so use one of them
for the metadata network and use the
second one for your local area network
and as I mentioned you can then share
the data using common networking
protocols to as many clients as you want
over your local neck over your local
network now these machines to be clear
the bottom ones the desktops that are
shown here do not have to be running X
and you do not need an X and license for
them of course with Mac OS 10 server you
get an unlimited client license you can
connect up as many clients to do file
sharing with the unlimited
licensing as you want and we don't care
whether that data comes from if it comes
from the sand that's great so as I
mentioned dual Gigabit Ethernet in the
excerpt g5 to support this now you could
also in another sense use this to backup
some kind of backup and use this as the
storage pool for a web service as well
whether it's a website QuickTime
streaming server or something like that
and have Internet clients connected into
them so but in any case there's two
networks there's the main network where
the data transfer happens and there's
the metadata network which is where the
file system does all of its operations
with the exception of doing the actual
block level access to storage which
happens over the fibre channel network
so when we deploy X in a kind of a post
production environment we deploy X in
two desktops I'm sure those of you in
the audience who do a lot of work in
video so you create a sand volume using
multiple xserve raids fibre channel
switch metadata controller server makes
a nice metadata controller and then you
set up the desktops with access to the
sand you can also on the same network
put a server there have it we share some
of that information using the QuickTime
streaming server so you can imagine a
workflow where you're doing video
editing and when for example at the end
of the day or at the end of the their
their time working on this project and
editor makes a QuickTime movie and posts
it into a special folder where the
QuickTime streaming server can share it
out to say producer or director who
wants to just check in over the network
without having to come down to the
editing Bay and of course you could put
appropriate security and access control
around all of that so this is an example
of using X and post-production workflow
another thing I'm not showing here that
you do have to setup again is that
private Ethernet network that connects
all of these systems and the metadata
controller
so there'll be a private Ethernet
network you can multi-home these systems
so you could use a second Ethernet
connection you can put another an
external well in a gigabit card into the
Xserve g5 for example or you could even
use the airport connection on the Xserve
g5 to connect it to the internet or
connect it to your local network but you
want to reserve one Ethernet port for
the metadata network computational
clusters are another area that we're
looking at with X an and in this case
we're showing a very large cluster for
small clusters as I said up to 64
systems can it connect to the sand so if
you had a small cluster of 4 16 32
systems you could actually have them all
connected directly to the sand over
fibre channel with block level access to
a single storage pool which means they
all can be using the exact same data
writing back to the exact same location
at the same time now if you have a very
large deployment and of course we use
the Virginia Tech one all the time 1100
g5s now they of course xn isn't
available so they weren't able to do
this but if they had chosen to do so
they could have actually created an
environment where of those 1100 say 20
of them 25 of them pick 55 of them say
55 of them were connected into the sand
directly and those 55 were the head
nodes for another 20 machines in the
cluster so the head node connected
block-level fibre channel directly into
the sand and then what it does is uses
NFS to share the data or AFP to share
the data to the other X serves in the
cluster the other 24 which it's
responsible now you've just and you
create a network between those 20 in the
head node another 20 in the head node
and you know now you've covered 1100
machines
and 455 of them their feeding data or
their getting data directly from the
sand at very high speed and they're able
to reshare that over say Gigabit
Ethernet to another 20 now I can tell
you that beach trying to reshare that
data over Gigabit Ethernet to another
1100 so this is a way of more easily or
more quickly distributing the data
around a computational cluster so I'm
gonna go through a few of our products
here and I'm run through some of that as
soon as I get another drink of water
here I'll run through some of this and
give you an idea of the kind of products
we're offering in the space and then I
want to give you an idea of why we think
we can be very successful in this space
in both video regular storage
consolidation projects that I've shown
you and also in computational clustering
where price performance is really what
counts so Mac OS 10 server qualified
with X with with X and delivers all
sorts of great services so most of you
caught the the earlier sessions on what
the next version of the server is all
about will be qualifying Tiger server
with xn as well when Tiger server is
closer to availability excerpt g5
phenomenal performance and a 1u server
an amazing box and I'm sure you've seen
some of the sessions on that as well we
do offer a fiber channel PCI card many
of you may not have realized that this
is a PCI card that goes into an X serve
or into it Power Mac and n delivers a
dual port 2 gigabits per second on each
channel so it's a really really nice
card it says 64-bit 66 megahertz PCI
card neat thing is we actually include
the cables and and these cables on the
street would be something like 150 200
dollars a piece but we include the
cables as part of the package for the
PCI card we have the Xserve raid this is
a three and
half terabyte solution in a 3u enclosure
again very very high speed if you stripe
across the channels as it says over 336
megabytes per second striping across the
two controllers very easy to set up this
is raid setup in three easy steps
very very nice solution this one's
actually qualified not only from Mac OS
10 Windows and Linux and as we said
it'll be qualified for X N and nice
solution $3.00 per gigabyte I have this
joke I have with just as a digression
with the guy who runs our server and
storage hardware I said he should only
have three slides on xserve raid the
first one should say $3.00 per gigabyte
the second one should say $3.00 per
gigabyte and the third one should say
did I mention $3.00 per gigabyte this is
a beautiful beautiful box an X and and
as I mentioned xn concurrent access
shared volumes up to 64 computers a lot
of great features in there from a high
availability or H a standpoint gives you
direct block level high speed access
over fibre channel to the storage
provide some really really cool volume
management capabilities comes with some
great remote administration tools and of
course we have great cross-platform
support with full interoperability with
the store next file system from Adi C
it's a really really complete solution
there and we have some great servers and
storage support options and we got a lot
of questions on this yesterday and I
just invite you to check out the latest
offerings that we offer in that space I
do want to run you through the prices
the thing that's key here and the reason
I'm running through some of the
marketing slides right now is that a lot
of people have said to us look we look
at I scuzzy we look at fibre channel we
not sure what's going to happen but it
sure looks like I scuzzy is going to
take over and when we look at it we say
well first of all you'll never be doing
video over ice Guzzi or not in the
foreseeable future and in fact it's very
very difficult to get any kind of high
performance out of ice Guzzi and when
you do you get into using really
interesting
proprietary hardware to do that and that
drives the expense way up
most people want I scuzzy because it's
cheaper than fiber channel that's the
clear reason why they want that sort of
thing
so we took a look at it we feel that
fiber channel is actually ripe for for
price reduction and so we're going to
change this market entirely and so we're
offering these products at an incredibly
low price compared to anything else on
the market yes I know
so 999 dollars for 10 server unlimited
client license xserve g5 $3,000 10
server is included with that the PCI
card is $500 with the cables included
the Xserve raid that $11,000 price is
with everything you can put in it
three-and-a-half terabytes of storage
hundred twenty eight Meg's of cash etc
etc and then xn under a thousand dollars
per node and unlimited capacity so so
some of you know that's that's easily
from a software standpoint that's a
fifth to a tenth of the price of
comparable solutions on the market today
and that's no exaggeration you can check
it out with other vendors it's a fifth
to tenth the price of comparable
solutions and many solutions actually
charge you on the basis of the amount of
storage that you put in place many
solutions charge you extra for the
metadata controller software all of that
stuff is built into X and so as I said
we're driving down the cost of this
infrastructure xn will be available in
the fall of 2004 so we're gonna have it
out to you in the fall of 2004 xn is in
beta today we made our first beta
available at NAB there is a program up
on the site Apple comm /x an and there's
a little link hidden over in the corner
referring to the beta program so you can
go to apple.com / x + / beta if you want
to sign up for the beta program now the
great thing about this in a sense is
we've had several hundred p
we'll already sign up for the program in
a couple of months that we've been
making it available and I will tell you
if you want to be part of the program
this is why I'm fascinated that there's
been that many people sign up you have
to have an extra raid you have to have a
qualified fiber channel switch and you
have to have two or more of either
Xserve g5s Xserve g-force or Power Mac
g5 and you have to have our fiber
channel PCI cards so in reality that's
that's you know 30 40 thousand dollars
worth of gear that you have to take
offline and use it for X and testing
yeah they really appreciate I know
there's a lot of you out there in the
audience today who are doing that with
us and we really really appreciate it if
you are interested in doing that go sign
up at the website and and we'll get you
the bit we are backlogged today in terms
of requests we've just been overwhelmed
with them but we really appreciate