Does LTSP actually work? [LWN.net]

Does LTSP actually work?

Posted Sep 29, 2006 18:43 UTC (Fri) by kbob (guest, #1770)
Parent article: The future of the Linux Terminal Server Project (Linux.com)

Does the LTSP model actually scale?

The claim is that you can use low-powered computers for the
workstations and run everything on a single high-powered server. Do
the economics of that work? It seems to me that they don't.

If 20 users are sharing 1 computer, then the shared system needs to be
much more powerful than a system that supports one user. Sure, it can
get by with less than 20X the CPU because no user uses the CPU 100% of
the time. But it needs the full complement of 20X the RAM and nearly
20X the disk space.

Further, screen rendering is one of the most demanding jobs a CPU
performs, so if you skimp on the workstations, the user experience
suffers. So can you really skimp on the workstation hardware?
Disk, yes. RAM, yes. CPU/video/net/buses? I doubt it.

Computer prices scale superlinearly with performance. A dual Xeon box
might cost 10X what a low-end Celeron/Sempron costs, but it certainly
doesn't provide 10X the system performance. Maybe 2X, depending on
workload.

Lots of people build LTSP systems using hand-me-down computers for the
workstations. But do those five year old workstations really run any
faster than they would with local disks and a RAM upgrade? You can
buy a lot of upgrades for the price of the server.

So I don't see it. LTSP looks more expensive than an
equivalent-performance set of standalone workstations. Easier to
administrate, sure, but more expensive for the hardware.

What am I missing?

I'd love to hear from someone who actually uses LTSP himself, as
opposed to administrating a cluster for a bunch of undergraduates/
secretaries/call center people or other people who're expected
to tolerate a low-quality computing experience.

kbob

(Log in to post comments)

Does LTSP actually work?

Posted Sep 29, 2006 22:44 UTC (Fri) by amikins (guest, #451) [Link]

LTSP can certainly be cost effective.

The first thing to note is that there is nothing that requires the actual 'server' portion to be a single monolithic machine. It can be a cluster itself, which can yield more efficient scaling of processing capacity versus cost. It gets trickier to ensure that task load is distributed correctly, but it's quite feasible.

The second is the utilization of local resources.. In many settings where a terminal server style approach makes sense, very few of the users are actually making use of most of their resources for consistently long periods of time. If every single user needed 100% of their ram, and most of their hard disk space potential, then yes, you'd need to scale the server accordingly. In practice, enough users are going to be underutilizing their portion of the resources that pooling them together is going to be a net win. Per user, you won't need as much ram and hard disk space as you would locally, both because of the above factors, and because of duplicated resources.

Additionally, there's data integrity issue. If all data is stored in one central location, extensible redundant storage becomes more managable.. And it's amazing just how much disk space really goes unused in a lot of 'low end' workstations.

Does LTSP actually work?

Posted Oct 5, 2006 23:05 UTC (Thu) by hazelsct (guest, #3659) [Link]

On modern hardware (within the past 5 years), the load of 20 users running things like evolution and firefox is not hard to bear.

Yes, you do need more RAM, but nowhere near 20X. Keep in mind that your RAM calculation ignores overlap in application binaries, shared libraries, and memory caches of data files, icons, etc., each of which needs just one instance in memory to cover all twenty users.

Does the LTSP model actually scale?

Posted Oct 8, 2006 7:59 UTC (Sun) by anton (subscriber, #25547) [Link]

Yes. We have been using X-terminals (not LTSP-based, but that does
not matter for answering this question) for 14 years. At the moment
we have about 25 X-Terminals for students, and right before deadlines
they all connect to one of the following two single-CPU machines: a
2.26GHz Pentium 4 with 2GB of RAM, or an 800MHz Alpha with 1GB of RAM
(slow by todays standards). In addition to the students on the
X-terminals, up to 60 students log into the machines from their homes
via ssh. The machines handle the load just fine, without significant
degradation of the user experience.

However, it depends on which software the users are using. Our
students mostly run an editor and compile and test their relatively
small programs now and then; no fancy Gnome and KDE applications. If
they used memory hogs, we would have to upgrade the RAM on the
machines; e.g., we upgraded the Pentium 4 from 1GB to 2GB when we
expected them to use Eclipse.

Most of our staff is also using X-terminals, using a wider range of
software (e.g., Mozilla, OpenOffice), but we don't concentrate as much
on one machine as the students.

There are some things that you just cannot do across the X-terminals,
like play some modern 3D game (although we did use the ACM flight
simulator with X-terminals a decade ago); not sure what else cannot be
done, because I never hit these limits.