is the process number of the yod command for the job,
obtained via ps. When yod receives this signal, it terminates the
parallel job and shuts itself down in an orderly fashion. Note that
from an interactive session, ctrl-C is interpreted properly.
What debugging tools are available on Cplant?
There currently is no source level debugger available for general
use. Debugging on the compute nodes via gdb is possible, but
requires a special configuration which has to be set up by Cplant
support.
What programming models are available on Cplant?
Cplant is a distributed-memory, MIMD machine which supports explicit
parallel programming.
In explicit parallel programming, the code developer must explicitly
decompose data structures into sub units and distribute them among the
nodes of the machine. The code written to execute on each node uses
standard languages (e.g., Fortran 77, C) for local processing.
Messages are passed between nodes using a message-passing protocol
(MPI or Portals) to coordinate processing.
What message passing protocols are available on Cplant?
The supported protocols for message passing on Cplant are Portals
and MPI. Portals provide the low-level communication facility
upon which all Cplant applications (user and support) are built.
Portals are essentially shared data structures between the
application process and the kernel which tell the kernel where to
deposit incoming messages.
Portals are available to the application through a user-level Portal
library, a system-level Portal library, or through direct
manipulation of Portal data structures. No special flags are
required to compile a code which uses Portals.
What libraries are available on Cplant?
The libraries which will be supported on Cplant are
o libc.a
o libm.a
o libdxml.a (DEC Extended math library)
o libmpi.a
Can I use sockets from within an application?
Sockets are not currently supported from within a compute node
application. Currently the only way to get data from the compute
partition to the outside world is through the file system.
In the near future, compute node applications will have the ability
to communicate with the service partition (and other compute node
applications) through an MPI-2 interface. Once this happens, it
will be possible to have a service node application receive messages
from the compute node application and funnel these messages off of
the machine through TCP/IP. In the distant future, there will be a
network partition which will also provide this capability, although
its exact functionality has not been determined yet.
What options are available for I/O from parallel applications on
Cplant?
A UNIX file system is currently the only file system avaialable.
All file I/O currently goes through yod. A parallel file system
is being developed.
What tools for performance analysis are available on Cplant?
None currently.
What tools for resource management are available on Cplant?
None currently.
What does a service node look like?
The hardware and operating system on a service node are identical to
that of the compute nodes. The service nodes have an extra ethernet
interface which allows them to be accessible to the LAN. The
service nodes also NFS mount user home directories from the central file
server.
What does an I/O node look like?
The I/O nodes are Alpha Server 1200's running Linux.
How much disk is available on Cplant now?
Available disk is system dependent.
Is HiPPI available for this system?
HiPPI is not available for Cplant.
Is there any way to get information about what is going on in my
program without instrumenting my code by hand?
Not currently.
What is the clock rate on the PCI bus?
The clock rate of the PCI bus is 33 MHZ.
What is the hardware latency for sending
messages? What is the software latency?
Check the Cplant
performance page.
How large is the message header for interprocessor messages?
The message header requires 64 bytes for all messages.
What binary format is used for data storage in files?
The DEC Alpha uses the little endian format.
Can message passing be used to communicate with processors external
to Cplant?
MPI is supported as a message-passing library on the compute partition
of Cplant, but off-machine connections for applications running on the
compute partition are not currently supported, but are planned.
What can I do to improve message passing performance?
Make sure the buffer you are sending from or receiving into is 4 or
8 byte aligned, i.e. the two least significant bits
(LSB) of the buffer address are 0.
Make sure your message length is a multiple of 4.
Send large messages to amortize message startup overhead.
Are there any known problems with the message passing?
The low-level Myrinet driver cannot currently send 1 and 2 byte
messages that do not start or end on a 4 byte boundary.
The Myrinet hardware has a built-in 4 second timer that causes a
network reset, when a message transfer takes longer than that
time. While that is enough time to transfer all the physical
memory of our nodes, it can cause problems in case of network
contention. For example, when all the nodes in an application
send large messages to a single node at the same time, the last
few messages to arrive at the single node may have been in the
network longer than the maximum of 4 seconds. We are working on
solutions to recover more gracefully from such a reset.
Where did Cplant get its name?
The Computational Plant derived its name from two of the main
concepts behind its inception. First, Cplant is a plant is the
sense of a power plant. Cplant provides compute cycles in much
the same way that a power plant provides electricity. Also, Cplant
is a living entity that will grow and be pruned on a three year cycle.
Each year new hardware will be added while older hardware will be
removed.
