[gPXE] I need a little ram disk after gPXE boot and before AoE sanboot

Sat Jan 30 10:34:29 EST 2010

Hi to everyone !

Please, i need some help with my project, I want to install windows XP on a server block device & AoE sanboot, diskless clients from it...preferably there'll be a single block device on a ram disk for booting multiple clients but i've been told that clients' writes would corrrupt the file system so, i need a way to make it read-only & would also need your help on the issues discussed below:

  Shao Miller wrote:  

  With that much RAM, you
  could fit a stripped-down XP entirely as a RAM disk.  I have a 204 MB XP
  which includes enough features that it has been extremely useful for
  diagnostic purposes.  With the right control of BOOT.INI, it boots on
  any hardware.  It has also been quite useful to gather up
  platform-specific device device nodes in the Device Manager and allow for
  easy export and pre-installation into a Windows image which is intended to be
  deployed to that platform. 

  It'd be interesting to
  strip down the size of win XP to some point only if it doesn't compromise the
  compatibility with the software i need to install & can be easily
  reproduced by others. 

  Yes, I suppose a temporary
  directory on a RAM disk could offer some performance benefits, as long as you
  don't run out of room!  Really, I would like to see something akin to
  Linux' rootfs on Windows, and will be working towards that at some point. 
  In a rootfs, there is no block device for the filesystem to be on, just as a
  network filesystem does not have a block device (as far as the client is
  concerned).  In this scenario, we'd merely use RAM without the block
  device layer, as the filesystem could be sized dynamically.  Anyway,
  that's not quite what you were talking about. :)  What RAM disk software
  do you think you'd be using?  I'm curious. 

  Sorry, but I'm
  not sure I fully understand this rootfs trick + for Linux there's already
  "fast diskless deployment framework"(faddef) (it's in French) &
  seems to be simple & portable...the RAM disk software I intend to install
  is the AR RAM Disk from ARSoft because it's free http://www.arsoft-online.de/index.php?option=com_remository&Itemid=43&func=fileinfo&id=3 

  This is not necessarily
  so.  You can use a local HDD for paging without having the filesystem
  available for general use (for saving files to).  Bo Brantén has offered
  SwapFs, for example.  Essentially, if you can make a filesystem
  available to the paging system but not to the user, you're pretty safe. 
  Also, the filesystem could be wiped and rebuilt easily at every boot.  I
  guess that would require some invetigation, though.  If you wish to be
  diskless, you wish to be diskless. 

  This is interesting
  but, Yes, i prefer to go diskless + a local disk makes the boot process last
  longer & if you add the time of rebuilding the FS at each boot it'll be
  inacceptable...I had an experience with a spyware that just automatically
  found any drive on the computer & copied itself on it, it copied itself
  even INSIDE the archives. (zip or rar or both i can't recall) 

  The HDDs on the
  clients?  Do you mean their SAN connection?  You said they should
  be diskless, so I assume you mean their SAN connection. 

  Well, I mean the HDDs that are in use on the PCs
  now, their avg throughput is around 50MB/s, if they are replaced by an
  ethernet link capable of ~70MB/s & the SSD on the server is capable of
  200MB/s max write, with 100 clients will it make any I/O jam on the server ?
  (the server should be connected by a gigabit ethernet link to the switch or
  more with link agregation)

  If the first client sends his 70MB
  in the first seconde then, the 100th client will be able to send its 70MB
  100sec after the first one :( resulting in a 0.7MB/s real network throughput,
  i'm I right ?

  I've already booted XP over a
  fast ethernet link ~10MB/s...to have the same performance i'll have to
  agregate 8, gigabit links...OUTCH ! 

  Nevermind,I'll try to post to gPXE in a few days
  when the discussion about the new release will settle a bit, I think i'll end
  up dividing the clients by department & have around 20 clients per
  server, for now it's better & cheaper than building a 10GB/s switched
  back bone & using a 10GB/s capable server. 

  Caching and a good
  copy-on-write strategy would be strong variables if your SAN target software
  can take advantage of the large quantity of RAM, in my opinion.  Using
  SSD for 100 clients would suggest a rather large storage capacity and a
  rather large expense, though...  Are you sure that's the setup you'd
  like to use? 

  Well, not quite sure !
  I'm doing my project(6months practice for my ACS) at a big company so, I
  don't want to tell them it's imposible, i'd rather say this is what it costs
  & here are the benefits! If the whole thing won't cost more than
  ~4000euro(~price of 100HDD) it'll be "sellable" & I'd have
  reached my goal, i've presented my project & said it was for ~10PCs, they
  let me start & then ask me how can we benefit from it?...we have 100 PCs
  on the same LAN. 

  If each of your 100 clients
  are _writing_ at that rate on average, I don't think you'd have capacity,
  since for 100 clients, your server would be requested to write roughly 5
  GB/s!  With iSCSI, the SAN software should report an I/O error to
  clients and they would retry the write.  Your user experience might be
  impacted.  An example might be if you had the media arts classes at a
  few schools simultaneously trying to save their work at the end of
  class-time.  There are some pretty quick SSDs out there, though. 
  What do you think? 

  What I think is, send
  me a link to those quick SSDs, how fast they are ? how much do they cost
  & where to get'em from ? 

  The director of the IT department
  wants to implement this for his engineers who are mainly using Oracle Client
  9i, PL/SQL Developper,Oracle Développer 6i & MS Office 2003.

  I've presented my project as
  intended for "small labs" in schools ...etc & they want me to
  make it an industry level beast  

  Alternatively, you could
  avoid the block device layer and simply use a network filesystem protocol for
  the user data.  Same possible integrity benefits.  Having a
  dedicated SAN per user makes for an easy quota enforcement policy (they get
  what they get due to the block device's size), but that would mean dividing
  up the backing store before-hand.  You could take a performance hit and
  use sparse files as the backing store, or some other layer that keeps a map
  of written blocks so that only used blocks are required on the backing store (your
  server's SSD). 

  Sorry, I have no experience with NFS, but why not if
  it's easy to back up & adds I/O performance ? I thought they could tape
  backup iSCSI block devices on a per PC basis (PCname-date), if it is as
  simple with NFS it's ok for me. 

  You want a sample WireShark
  I/O graph of TCP/IP over the course of a day?  Again, someone on the
  gPXE mailing-list might have one readily available for you.  Feel free
  to ask. :) 

  Yea, it should be cool...as long as it doesn't
  provide ip adresses or network packets information :) just tell me a bit
  about the number of clients & the network topology that produced the
  graph. 

  I am not familiar with device-mapper
  for FreeBSD.  The driver for NetBSD might be portable.

  Without copy-on-write, I have compiled the AoE
  'vblade' service on FreeBSD with success.  Did you mean your server
  would be an AoE or an iSCSI server?  Or both? 

  At the begining the server was meant to run AoE
  & iSCSI for 10 to 14 clients but now with a hundred clients i think i'll
  have to separate the roles at some point, the number of clients per
  department is between 7 & 29 (avg of 18 on 5 departments) but all 102
  clients are on the same LAN.

  Me too, i've successfully
  compiled & installed vblade on freebsd 6.4 & did an AoE boot from it
  (at home) but after the user login the desktop wallpaper shows & the
  "server" freezes, always.(it's less powerful P4 at 1.6GHz than the
  client & has less RAM, 128MB & 1GB RAM on the clientP4 at 2.66GHz) 

  Hmm...  If most I/O
  are reads and the server has good caching and lots of RAM, it mightn't be
  impossible.  When clients boot and are using the same applications, much
  of this could be cached into the server's RAM.  With 8 GB of RAM, you
  could fit an entire Windows installation into RAM and have it ready to go for
  clients.  In my opinion, it's really the writes that are of primary
  concern.  Writes for paging, writes for user data.  A local HDD
  helps with this. 

  That's why i wanted to use a single block device
  hosted on a RAM disk, at the begining I thought I could make it read only
  with chmod but i couldn't, do you have any idea for making a block device
  read only ?  

  At the begining I wanted to use an enhanced write filter but it's only
  available for XPE, maybe windows SteadyState will do the same job ?. 

  What are the needs for your
  environment?  Do clients share a lot of data, such as OS, drivers,
  applications?  How much writing will they be doing?  Would a network
  filesystem make more sense for user data writes? 

  On top of windows XP the following softs are
  installed: Oracle Client 9i, PL/SQL Developper,Oracle Développer 6i & MS
  Office 2003.

  As i've been told by my tutor,
  the clients don't generate a lot of data locally & they are mainly used
  to connect to & administer databases on the DB servers that are in the
 Data Center . 

  Thank you so much for
  all you efforts & Take care of yourself !

  The Mad One. 

Thanks to everyone !The Mad One.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://etherboot.org/pipermail/gpxe/attachments/20100130/a530c86a/attachment-0001.html 