[Author Prev][Author Next][Thread Prev][Thread Next][Author Index][Thread Index]

Re: [school-discuss] linux distributions for low resource computers,



I'm adding another piece to this puzzle:

power and cooling are an issue that can make or break this equipment.
For colder climates than Atlanta, extra heat making devices in each
classroom is not so much of an issue provided the classroom has the
power availability to support it. For warmer climates having as few heat
sources in the classroom as possible for August-October and again from
March-May is very important.

Several of the thin clients I've seen are nearly toasters while a
wonderful few are safe for little fingers to touch. The more horsepower
the client has, the more power it consumes and the more heat it gives
off (with other things being held equal).

The servers, however, have to make up the difference. The diskless
workstation server is mostly network and disk space so it will sit
somewhat idle from the power consumption standpoint since the cpu is the
main power point. The thin client servers will be mostly chugging along
at full speed during the day. By isolating the servers in a server
closet (enterprise model) supplemental power and cooling can be
localized as needed. With smaller servers distributed thought the
building, that aspect becomes more of a challenge but somewhat less of
an issue (up to a point). Older schools with flaky power may need to run
new circuits for computer use.

Here's where I'm getting challenged. As Linux thin clients progress
towards the diskless workstation model, we get closer to the standalone
desktop of Mac and Windows notoriety. The standalone desktop is a huge
power waster. It sits waiting for user input for the vast majority of
its time. The exceptions are the game rigs kids play with. I don't see
much use for those in schools (yet!). Thin clients work very efficiently
by keeping a server running with little to wasted cycles. The clients
are very small and generally working quite efficiently as well. For the
moment, the ideal is the "chubby" client (an LNS internal term that
describes a midpoint between thin client and diskless workstation) that
runs the system hogs locally as needed and all else on the server.

I expect to have to re-evaluate this shortly as there is some new
capabilities in the server pipelines that may make thin client the most
efficient AND capable of multimedia. This new stuff has the ability to
power cycle unused cpu cores on the fly based on load. So a 2 cpu, quad
core can be backed down to 1 core on each cpu for "normal" loads and
save lots of power and still crank up extra horsepower when needed. This
is why I like AMD cpu technology :-)

On Tue, 2008-06-03 at 19:54 -0400, Daniel Howard wrote:
> OK, so $300 for a diskless workstation vs. $150 for thin clients.  It's
> only 3 dB in cost increase I guess...and as long as the choice of 
> diskless doesn't reduce the number of classroom PCs you can afford; 
> we've found here in Atlanta that 2:1 is optimal given classroom 
> electricity and space limitations, and even our 1:1 fifth grade 
> classroom teachers have now asked to go back to 2:1 because they're 
> tired of resetting the circuit breaker and they feel that 2:1 is 
> adequate for the vast majority of their needs.
> 
> But I'd really like to know how many thin clients you can load up with 
> Java games, YouTube, and 3d accelerated programs *before* a server bogs 
> down, and as a function of the server specs (quad core single CPU all 
> the way to Jim's monster servers).  This could make a huge difference in 
> the overall cost of the two solutions.  If a currently available $500 
> classroom server can handle at least 10 clients doing graphically 
> intensive activities all at the same time, then it's a comparison of 
> $200 ($150 client plus $50/client server cost) per workstation to $300 
> per workstation.  Worse, in 5 years when it's time to upgrade the school 
> computers, you have to do it for all workstations, not just the servers. 
>   For our school with 400 workstations, that could be say $200/diskless 
> node (assuming cost comes down while performance goes up) or $80,000 vs. 
> replacing 40 servers at $500 each (no cost reduction, just performance 
> improvement) or $20,000.  Nope, I still think the thin client solution 
> wins, it's just that you need more servers to do graphically intensive 
> stuff.
> 
> Would like to read Robert A's post, but the link below doesn't work, can 
> you correct and resend?
> 
> Best,
> Daniel
> 
> 
> Dean Montgomery wrote:
> > On June 3, 2008, James P. Kinney III wrote:
> >> Can you provide some specs on the server you are using? My LTSP
> >> servers are pretty beefy monsters (2-4 cores, 2-4 CPU's, 8+GB RAM,
> >> multiple Gbit NICs, etc) that run 100-150 thin clients each.
> > 
> > That is in the ballpark that we have been using.  Spend money on the
> > server and save on the client.
> > 
> > With LTSP when we loaded up a lab of 30 students doing Java games,
> > youtube,  or trying to do any 3d accelerated programs the server and
> > the network would bog down.  Even powerpoint slide transitions are
> > choppy on LTSP.
> > 
> > With diskless once the java game (or google earth) is loaded, there
> > is no more network traffic.  A bonus is the kernel caches the program
> > in the client's ram so the next time openoffice or firefox is clicked
> > on, it starts almost instantly with very low network traffic.   We
> > can have 100+ kids and teachers doing a java game, youtube,
> > openoffice, or google earth without any performance issues.
> > 
> > 
> >> I would expect for diskless clients the key factors will be
> >> primarily network bandwidth followed by hard drive throughput.
> > 
> > That is why we do a gig connection from the server to the switch.
> > Hard drive is a RAID - kernel does some file caching so the drive
> > isn't thrashing.  Using sar I've noticed that the nightly backup is
> > the only time that the hard drive and CPU start pushing 100%.
> > Throughout the day they rarely reach 100%.
> > 
> >> The main reason I have been avoiding the diskless format is heat
> >> and the need for fans. A silent environment is a big plus from the
> >> thin client format. Although I do have my hands on a new fanless
> >> client from VIA (pico-format) that can be either diskless or thin
> >> format.
> > 
> > Yes I agree.   We have been focusing on cost savings instead of noise
> > reduction.  The computers are quieter than fat clients but they still
> > have a CPU and Power fan.  AMD's Cool-n-quiet throttles it down even
> > more.
> > 
> > 
> > Here is a screenshot of a diskless desktop... 
> > http://www.sd73.bc.ca/misc/linux/sd73.png ... and yes all the
> > students can do this at the same time with no slow-downs or lags. : )
> > Think about it... the movie file is travelling over the network as
> > compressed avi and then is decrypted/displayed on the client
> > workstation.  Beryl is loaded into client RAM and doesn't require
> > much network IO.
> > 
> > Here is a screenshot from on the server running about 150 diskless
> > clients.  Notice the gkrellm monitor on the left showing 46% & 10%
> > cpu usage; bond0 is 2 bonded NICS.  MEM is hardly being used. 
> > http://www.sd73.bc.ca/misc/linux/skss_charts.png
> > 
> > The rest of the screen shots are from sar and they show that the
> > biggest load on the server is during the nightly backups. 
> > http://www.sd73.bc.ca/misc/linux/
> > 
> > ==== If you still don't believe me see what Robert Arkiletain says
> > about diskless.  (Robert wrote Teacher Tool for LTSP.)
> > 
> > <htp://groups.google.ca/group/bcfosss/browse_thread/thread/f4b51edeffdc1e56?hl=en>
> > 
> > 
> > 
> 
> -- 
> Daniel Howard
> President and CEO
> Georgia Open Source Education Foundation
> 
-- 
James P. Kinney III          
CEO & Director of Engineering 
Local Net Solutions,LLC                           
http://www.localnetsolutions.com

GPG ID: 829C6CA7 James P. Kinney III (M.S. Physics)
<jkinney@xxxxxxxxxxxxxxxxxxxxx>
Fingerprint = 3C9E 6366 54FC A3FE BA4D 0659 6190 ADC3 829C 6CA7


-- 
This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.