Re: [VirtualGL-Users] Single application runs very slow withVirtualGL/TurboVNC while others run fine

[DRC <dco...@us...>]

> My Questions are:
> 
> - Is there a way to further debug a VirtualGL session?

You may be encountering a known issue with ANSYS and VGL, described in
"Notes for ANSYS and Pro/E" in the following section of the VGL User's
Guide:
http://www.virtualgl.org/vgldoc/2_1final/#hd0015
If this is indeed the cause of your slow-down, then you should be able to
work around the issue by setting the environment variable VGL_INTERFRAME=0
on the server (if you are displaying using the VGL Image Transport) or
setting the environment variable VGL_FPS=20 on the server (if you are
displaying using TurboVNC.)

If this doesn't work, then the next step would be to run VirtualGL in
profiling mode:

vglrun +pr {your application}

You can also setenv VGL_PROFILE=1 on the client machine to get similar
output from vglclient.  Next, load your application and (if possible) get it
to automatically rotate a model without mouse interaction.  This should
produce a steady image stream into VirtualGL, so you can monitor the
profiling output on both server and client and (hopefully) figure out where
the bottleneck is.


> - Are there any general tipps to further increase overall performance?

If other applications are performing adequately, I suspect this is
application-specific rather than a systemic issue.  Rarely, an application
will encounter issues with a remote X display because the application uses
the X protocol inefficiently.  In these cases, the application may perform
better in TurboVNC, since TurboVNC acts like a local X display.  In the case
of ANSYS, the app is using OpenGL inefficiently.  It can send thousands of
back-to-back glFlush() calls, which works OK on a local display but tends to
trip up VirtualGL's pixel readback heuristics.


> - Is there any practical experience how many users may use VirtualGL 
> simultanously?

Depends on the application and the hardware capabilities.  We've been able
to put as many as 4 per GPU with "typical" CAD applications without
encountering any severe performance lag, but our studies on this topic are
far from incomplete.


> - Is there a way to utilize multiple cpu cores, as rotating 
> an object seems to 
> produce 100% cpu usage on one core for a short period of time 
> (ssh process)?

VirtualGL uses a separate thread to perform image compression, so if the
application is using one CPU to do its computation, VirtualGL shouldn't
theoretically interfere with it.  You can try enabling multi-threaded
compression (vglrun -np {number_of_threads}), but it's been our experience
that this doesn't really help all that much unless the server has a bunch of
old, slow CPUs and is displaying to a newer client.  Usually, the
client-side decompression is the main bottleneck when running on a LAN.


> - Is there also a commercial support for VirtualGL/TurboVNC available?

Absolutely!
http://www.sun.com/servers/cr/visualization/shared.jsp