[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[TIGGE #XLU-757491]: tigge ldm server specs



Hi Doug,

Mike is on travel at the moment, so I will jump in and try to answer your
question.

re:
> Disk I/O seems to be one of the larger demands when memory in
> inadequate.

Yes.  That is why we recommend significantly more memory than will
hold the size of the LDM queue.  On our idd.unidata.ucar.edu cluster
we typically run an 10-12 GB queue on machines that have 16 GB of
physical RAM.

> Currently I've got the queue set at 8GB, with 16 GB of physical
> memory available on our SUN system.

That should be adequate, but it will depend on what processing other
than the LDM is being done.

> Products (individual GRIB-2 messages) are processed out of the queue
> to local disk, then larger files are built
> from these pieces using the cat command from local disk to the SAN
> disk.  When the queue is set larger,
> pqact processing gets way behind.  Is this because swap memory needs
> to be used for the queue?

It is possible that you are seeing a swap-related issue when the queue
is set to be larger than 8 GB.  This would be my best guess as to what
is occurring _if_ all other parameters are kept fixed (meaning that
you did not change your processing) when you changed the size of your
LDM queue.

I am not sure if Mike (or someone else here including me) has commented to
you and other TIGGE folks that we migrated to use of a Web 100 kernel on
our Linux LDM/IDD platforms.  The Web 100 kernel provides for dynamic
management of TCP buffer sizes so that more space is allocated for
those connections that need it AND less is allocated for those that
do not.  This feature allows us to run our idd.unidata.ucar.edu cluster
nodes with significantly larger LDM queues than were possible using the
stock kernel in the Fedora Core releases.  Mike told me, however, that
he read that the Web 100 tuning capability has been built into the most
recent Fedora Core distributions, so switching to a Web 100-specific
kernel may no longer be necessary.  Please note that we can not comment
on if this is the case in distributions other than Fedora Core, so 
the comment about Web 100 tuning in SuSE (or Debian, Ubuntu, Slackware,
Knoppix, Gentoo, etc.) ditributions may not apply.

One thing I did not notice in Mike's system recommendations to you
was the need to purchase 64-bit systems.  This may seem obvious when talking
about 32 GB of system RAM and a 20 GB LDM queue, but I want to make sure
that we are all aware of this requirement.

Finally, we have been having discussions here at the UPC about the
impact of the number of processors that an LDM ingest system has.
While it might seem like more processors are better, there is an issue
with LDM queue write lock contention.  Our experience on our old Sun SPARC
IDD toplevel relay node, thelma.ucar.edu, hinted that upping the number
of processors was not necessarily a good thing.  Experience on that
machine also showed that it was advantageous to put "collector"
machines in "front" of the relay nodes so that they (the front-end machines)
made the requests for data from the upstream sources, and the back-ends
did the data relay.  We use two front-end machines to redundantly collect
the data from upstream IDD nodes and then feed the data in single connections
to each of our cluster back-end machines.  This cuts the number of processes
contending for write locks from about 17-20 down to 2.  This configuration
works very well for our cluster where the back-ends each can support
a large number (we have seen over 200 on each node with no introduction
of latency from overload) of connections to downstream machines.  Since the
connection between the "collectors" and "back-ends" is Gbps, there is
little, if any, need for large LDM queues on the collectors -- we typically
run the collector machines with 2-6 GB queues mainly since we want to be
able to bring up new nodes, or restart existing nodes after system upgrades,
etc. and have their queues populated with as close to an hour's worth of
data as possible.

The point I am trying to make in the above somewhat rambling narrative is that
we believe that a single machine approach for TIGGE may not scale as more sites
start sending ensemble data to the archive centers.  Our recommendations to
the CMA about using more than one node of their cluster for TIGGE collection
and dissemination activities is based on our own experience in creating our
IDD relay cluster and in our trying to predict the load that could/will be
added to the nodes running at the TIGGE archive centers.  It is my opinion
that single machine implementations are likely to not be sufficient for the
work needed.

We can/should discuss these issues at more length before you/ECMWF/CMA embarks
on the next round of equipment purchases.

Cheers,

Tom
****************************************************************************
Unidata User Support                                    UCAR Unidata Program
(303) 497-8642                                                 P.O. Box 3000
address@hidden                                   Boulder, CO 80307
----------------------------------------------------------------------------
Unidata HomePage                       http://www.unidata.ucar.edu
****************************************************************************


Ticket Details
===================
Ticket ID: XLU-757491
Department: Support IDD TIGGE
Priority: Normal
Status: Closed