I'll second everything that's been said already about memory caching and
performance.
We too have had, at times, many different users accessing the same set of model
data via NFS from a completely average performance server, over just 100Mbs
connections and never really noticed any performance issues. If there was a
measurable difference between 3 and 30 users, it was not noticeable from the
user perspective.
Typically, even though some of the model files may be large, the analysis tool
you are using is only pulling out much smaller subsets (i.e. specific levels,
times, parameters) from the larger file, generally without having to read the
whole file (which is certainly the case with GEMPAK files as they are indexed
to help the client find the specific subset of data it wants very efficiently)
Secondly, I believe the bottleneck is more likely going to be the speed at
which the client can numerically process the data it receives.
Third, if you could orchestrate it where all 30 users hit the Go button exactly
at the same moment, you might notice some slowdown, but our experience has been
that even though you may be having an instructor leading a class through the
same set of steps, the users follow along at significantly different paces and
the work is spread out sufficiently.
Have you actually experienced performance problems in the past or are just
planning ahead? In a different, but closely related situation recently, we
had 45 users in a class that met in a Windows-only computer lab who needed to
run Garp. They were going to all be running Garp on one Linux server and
displaying via an X11 client on Windows. As I found out about the need for
this only a couple days before the class started, I scrambled and quickly
commandeered a high-power server that had yet to be unboxed. I sat nervously
in the back of the class, logged in and watching the stats to see how the
performance would be the first time they tried this. To my surprise, the
server barely broke a sweat with 45 users running Garp remotely all at once.
After a month, I moved it all to a smaller server (so I could put the fast one
to its originally planned use) and had no problems. All the remote X11 display
data over the network to a lab on a far side of campus wasn't an issue either,
even if they were running animations.
Garp is super-efficient. (Too bad the same can't be said for IDV which is a
huge resource hog and much slower)
------------
David Wojtowicz, Sr. Research Programmer
Coordinator of IT Services, School of Earth, Society and Environment (SESE)
University of Illinois at Urbana-Champaign
davidw@xxxxxxxxxxxx +1 217 333 8390
> -----Original Message-----
> From: ldm-users-bounces@xxxxxxxxxxxxxxxx [mailto:ldm-users-
> bounces@xxxxxxxxxxxxxxxx] On Behalf Of Bret Whissel
> Sent: Tuesday, March 29, 2011 10:05 AM
> To: ldm-users@xxxxxxxxxxxxxxxx
> Subject: Re: [ldm-users] solid state drive(s) for ldm?
>
> If the server has sufficient RAM, then the first disk access will put the
> model
> data into kernel cache, and subsequent file access will read from RAM. It is
> more likely that the network connection will be the bottleneck. Of course,
> the first disk access may be faster with a SSD, and you have fewer concerns
> about media errors over the long haul. In contrast, if you have many
> different files being accessed (in which case the kernel disk cache doesn't
> provide much advantage), perhaps an SSD could improve throughput. (Be
> careful about the type of drive, however, as the NAND-type storage has a
> limited number of disk writes over the life of the drive.)
>
> (We run a classroom of 30 Linux workstations accessing NFS-mounted data
> from a NAS-type file server. We are generally pleased with performance.)
>
> Bret Whissel
> SysAdmin
> Earth, Ocean and Atmospheric Science Department Florida State University
>
> On Tue, 2011-03-29 at 09:43 -0500, Neil Smith wrote:
>
>
> Was wondering if anyone has considered or made use of speed
> advantages of solid state drives (SSD) for serving decoded ldm data to
> gempak, garp, and too-be AWIPS2 processes running on network clients?
>
>
>
> -- particularly in the classroom environment where visualization tools
> from 20+ network clients are hitting the same $GEMDATA/models/<model>
> collection at the same time.
>
>
>
> When would SSDs be worthwhile? If the (NFS) clients are on a 100
> Mbps subnet and server is on separate 1000 Mbps subnet, is the network
> the bottleneck, leaving modern drives or SSDs of negligible difference?
>
>
> -Neil
>
> ---
> Neil Smith neils@xxxxxxxx
> Comp. Sys. Mngr., Atmospheric Sciences
>
>
>
