| 1X |
An InfiniBand interface width. 1X defines an interface with two differential
pairs, one transmit, one receive. Provides 2.5 Gbit/s full-duplex connections. |
| 4X |
An InfiniBand interface width. 4X defines an interface with eight differential
pairs (four per direction), four transmit, four receive. Provides 10 Gbit/s full-duplex
connections. |
| 4X DDR |
Double data rate InfiniBand interface. 4X defines an interface with eight
differential pairs (four per direction). Providing 5.0 Gbits/s per differential
pair for 20 Gbit/s full-duplex. |
| 12X |
An InfiniBand interface width. 12X defines an interface with 24 differential
pairs (12 per direction), 12 transmit, 12 receive. Provides 30 Gbit/s full-duplex
connections. |
| 12X DDR |
Double data rate InfiniBand interface. 12X defines an interface with 24 differential
pairs (4 per direction). Providing 5.0 Gbits/s per differential pair for 60 Gbit/s
full-duplex. |
| AMPI |
Adaptive MPI. An MPI implementation developed at the University of Illinois
that can exploit virtual processor techniques. |
| API |
Application programmer's interface. Syntax and semantics for invoking services
from within an executing application. All APIs shall be available to both Fortran
and C programs, although implementation issues (such as whether the Fortran routines
are simply wrappers for calling C routines) are up to the supplier. |
| ARMCI |
Aggregate remote memory copy interface. A one-sided communication library
that provides an extensive set of RMA. See http://www.emsl.pnl.gov/docs/parsoft/armci/ |
| BLAB |
Aggregate Bidirectional Link Bandwidth. BLAB is defined as the minimum of
the aggregate memory bandwidth, aggregate bus bandwidth, or the sum bidirectional
link peak user payload data bandwidth. |
| blocking operation |
An operation that does not complete until the operation either succeeds or
fails. For example, a blocking receive will not return until a message is received
or until the channel is closed and no further messages can be received. |
BlueGene/L
BG/L |
A jointly funded research partnership between IBM and the Lawrence Livermore
National Laboratory as part of the U.S. Department of Energy ASC Advanced Architecture
Research Program. Application performance and scaling studies have recentlybeen
initiated with partners at a number of academic and government institutions, including
the San Diego Supercomputer Center and the California Institute of Technology.
This massively parallel system of 65,536 nodes is based on a new architecture
that exploits system-on-a-chip technology to deliver target peak processing power
of 360 teraFLOPS (trillion floating-point operations per second). The machine
is scheduled to be operational in the 2004-2005 time frame at price/performance
and power consumption/performance targets unobtainable withconventional architectures.
See http://www.research.ibm.com/bluegene/ |
| broadcast operation |
A communication operation in which one processor sends (or broadcasts) a message
to all other processors. |
| buffer |
A portion of storage used to hold input or output data temporarily. |
| Charm++ |
A parallel C++ library developed at the University of Illinois. |
| Clos |
A network topology named after its inventor Charles Clos. |
| cluster |
A set of SMPs connected via a scalable network technology. The network shall
support high bandwidth, low latency message passing. It shall also support remote
memory referencing. |
| collective communication |
A communication operation that involves more than two processes or tasks.
Broadcasts, reductions, and the MPI_Allreduce subroutine are all examples of collective
communication operations. All tasks in a communicator must participate. |
| communicator |
An MPI object that describes the communication context and an associated group
of processes. |
| critical path |
The serial chain of dependencies that most limits forward progress. |
| DDR |
Double data rate |
| device driver |
Software to function the host channel adapter devices on a node. |
| DHCP |
Dynamic host configuration protocol. DHCP enables individual computers on
an IP network to extract their configurations from a server (the 'DHCP server')
or servers, in particular, servers that have no exact information about the individual
computers until they request the information. DHCP is often used to reduce the
work necessary to administer a large network (e.g., managing IP addresses). |
| DPCL |
Dynamic Probe Class Library |
| fairness |
A policy in which tasks, threads, or processes must be allowed eventual access
to a resource for which they are competing. For example, if multiple threads are
simultaneously seeking a lock, no set of circumstances can cause any thread to
wait indefinitely for access to the lock. |
| FT |
Fault tolerance or fault tolerant |
| GiB |
gibibyte. Gibibyte is a billion base 2 bytes. This is typically used in terms
of random access memory and is 230 (or 1,073,741,824) bytes. |
| GB |
gigabyte. Gigabyte is a billion base 10 bytes. This is typically used in every
context except for random access memory size and is 109 (or 1,000,000,000)
bytes. |
| GPL |
General Public License. A legal software license arrangement developed by
GNU to promote open software. The licenses for most software are designed to prevent
users from sharing or changing it. By contrast, the GNU General Public License
is intended to guarantee the freedom to share and change free software to ensure
the software is free for all its users. The GPL is designed to make sure that
anyone can distribute copies of free software (and charge for this service if
they wish); that they receive source code or can get it if they want; that they
can change the software or use pieces of it in new free programs; and that they
know they can do these things. The GPL forbids anyone to deny others these rights
or to ask them to surrender the rights. These restrictions translate to certain
responsibilities for those who distribute copies of the software or modify it. |
| GUI |
Graphical user interface. A type of computer interface consisting of a visual
metaphor of a real-world scene, often of a desktop. Within that scene are icons,
which represent actual objects, that the user can access and manipulate with a
pointing device. |
| HCA |
Host channel adapter. IBA expansion card that interfaces the IBA interconnect
to the cluster node I/O subsystem. |
| HEC |
High-end computing |
| hot spot |
A memory location or synchronization resource for which multiple processors
compete excessively. This competition can cause a disproportionately large performance
degradation when one processor that seeks the resource blocks, preventing many
other processors from having it, thereby forcing them to become idle. |
| HPC ULPs |
High performance computing upper layer protocols. HPC ULPs include MPI, IPoIB,
SDP, and Sandia Portals. |
| HT |
HyperTransport is an I/O link. With clock speeds of up to 1.4 GHz and DDR
signaling, HyperTransport technology provides an effective throughput of 2.8 gigatransfers
per pin-pair on a 32-bit link. This results in a maximum aggregrate throughput
of 22.4 gigabytes per second, per link. (See http://www.hypertransport.org/tech/index.cfm) |
| IBA |
InfiniBand architecture |
| IBTA |
InfiniBand Trade Association (See http://www.infinibandta.org/ibta/) |
| InfiniBand access layer |
Includes the user-mode components for management services, SM query, connection
management, and work request processing, and the kernel mode components for InfiniBand
PnP, management services, resource management, connection management, work request
processing, and user-level proxy agent. |
| IPoIB |
Internet protocol over InfiniBand. IP specifies the format of packets (also
called datagrams) and the addressing scheme. |
| iSCSI |
Internet SCSI (Small Computer System Interface). An IP-based storage networking
standard for linking data storage facilities, developed by the Internet Engineering
Task Force (IETF). By carrying SCSI commands over IP networks, iSCSI is used to
facilitate data transfers over intranets and to manage storage over long distances. |
| iSER |
iSCSI extensions for RDMA |
| kDAPL |
Kernel Direct Access Programming Library defines a single set of kernel-level
APIs for all RDMA-capable transports. The kDAPL mission is to define a transport-independent
and platform standard set of APIs that exploits RDMA capabilities, such as those
present in IB, VI, and iWARP. |
| kernel modules |
Changes to the Linux kernel needed to support the rest of the OpenIB software
stack. |
| LAB |
Aggregate link bandwidth. LAB is defined as the minimum of the aggregate memory
bandwidth, aggregate bus bandwidth, or the sum of unidirectional link peak user
payload data bandwidth. |
| LAPI |
Low-level application programming interface. An active-message-type API for
optimal communication through the IBM SP switch. Provides reliable, unordered
communication between all processes in the MPI world. |
| latency |
The time interval between the instant at which an instruction control unit
initiates a call for data transmission, and the instant at which the actual transfer
of data (or receipt of data at the remote end) begins. Latency is related to the
hardware characteristics of the system and to the different layers of software
that are involved in initiating the task of packing and transmitting the data. |
| LVDS |
Low voltage differential signaling. An electrical spec (EIA-644) used by InfiniBand.
LVDS is designed with an output voltage swing of 350 mV at better then 400 Mbps
into a 100 ohm load, across a distance of about 10 meters. |
| MB |
Megabyte is a million base 106 bytes. This is typically used in
every context except for random access memory size and is 106 (or 1,000,000) bytes. |
| MiB |
Mebibyte is a million base 2 bytes. This is typically used in terms of Random
Access Memory and is 220 (or 1,048,576) bytes. |
| MPI |
Message passing interface. An industry standard, message-passing protocol
that typically uses a two-sided send-receive model to transfer messages between
processes. |
| MPI-2 |
Extensions to the MPI standard. |
| MPI I/O |
An MPI extension allowing for the manipulation of files on different file
systems. |
| MR |
Mandatory requirement. Mandatory requirements are items that are essential
to the University and reflect the minimum qualifications an offeror must meet
in order to have their proposal evaluated further for selection. |
| MTBF |
A measurement of the expected reliability of the system or component. The
MTBF figure can be developed as the result of intensive testing, based on actual
product experience, or predicted by analyzing known factors. |
| NIC |
Network interface card. An expansion board you insert into a computer so the
computer can be connected to a network. Most NICs are designed for a particular
type of network, protocol, and media, although some can serve multiple networks. |
| nonblocking operation |
An operation, such as sending or receiving a message, that returns immediately
whether or not the operation was completed. For example, a nonblocking receive
will not wait until a message is sent, but a blocking receive will wait. A nonblocking
receive will return a status value that indicates whether or not a message was
received. |
| open source |
Software products provided under an open source license(s) found at http://www.opensource.org. |
| parallelism |
The degree to which parts of a program may be concurrently executed. |
| PCI Express |
A dual-simplex, point-to-point serial differential low-voltage peripheral
component interconnect. Previously known as 3GIO and Arapahoe. PCI Express allows
a bandwidth up to 500 MB/s duplex for each link, 8 GB/s for sixteen lanes (x16). |
| PCI-X |
A follow-on initiative to PCI (peripheral component interconnect). PCI-X allows
a bandwidth up to 1 GB/s for 64 bit bus running at 133 MHz. [Note that we distinguish
PCI-X and PCI Express.] |
| PERUSE |
MPI performance examination and revealing unexposed state extension specification;
the specified API. |
| PMPI |
Profiling interface for MPI specified by the MPI standard. |
| Portals (Sandia Portals) |
Low-level API providing reliable and ordered communication for Lustre.( See
http://sourceforge.net/projects/sandiaportals.) |
| POSIX |
Portable Operating System Interface. A set of IEEE standards designed to provide
application portability between UNIX variants. IEEE 1003.1 defines a UNIX-like
operating system interface, IEEE 1003.2 defines the shell and utilities and IEEE
1003.4 defines real-time extensions. |
| QDR |
Quad data rate |
| RC |
Reliable connection |
| RDMA |
Remote direct memory access. RDMA capability allows processes executing on
one node of a cluster to be able to "directly" access (execute reads
or writes against) the memory of processes within the same user job executing
on a different node of the cluster. |
| reduction operation |
An operation, usually mathematical, that reduces a collection of data by one
or more dimensions. For example, the arithmetic SUM operation is a reduction operation
which reduces an array to a scalar value. Other reduction operations include MAXVAL
and MINVAL. |
| RHEL |
Red Hat Enterprise Linux |
| RMA |
Remote memory access. A user-level communication protocol that provides ability
for a task to access memory of another task by the use of put/get operations. |
| RTS |
Run-time system |
| Scalability |
Tested on 4,096 node physical fabrics and scaling properties simulated up
to 16,384 nodes. |
| SDP |
Sockets direct protocol. SDP is an IBA-specific protocol defined by the Software
Working Group (SWG) of the IBA. The SDP specification maintains traditional sockets
SOCK STREAM semantics as commonly implemented over TCP/IP, as well as support
for byte-streaming over a message passing protocol, including kernel bypass data
transfers and zero-copy data transfers. |
| SDSM |
Software-based distributed shared memory |
| SMP |
Shared memory multiprocessor. A set of CPUs sharing random access memory within
the same memory address space. The CPUs are connected via a high speed, low latency
mechanism to the set of hierarchical memory components. The memory hierarchy
consists of at least processor registers, cache and memory. The cache shall also
be hierarchical. If there are multiple caches, they shall be kept coherent automatically
by the hardware. The main memory may be a non-uniform memory access (NUMA) architecture.
The access mechanism to every memory element shall be the same from every processor.
More specifically, all memory operations are done with load/store instructions
issued by the CPU to move data to/from registers from/to the memory. A single
SMP may be partitioned into one or more nodes. |
| SOW |
Statement of work |
| SPMD |
Single program multiple data |
| synchronization |
The action of forcing certain points in the execution sequences of two or
more asynchronous procedures to coincide in time. |
| Test harness and modules |
Software to automatically test the functionality, performance, reliability,
and robustness of the components of the OpenIB software stack. |
| TF |
Teraflop. A measure of the peak computing power of a machine in 1012
floating point operations per second. |
| thread |
A single, separately dispatchable, unit of execution. There may be one or
more threads in a process, and each thread is executed by the operating system
concurrently. |
| TLP |
Thread level parallelism |
| UD |
Unreliable datagram |
| uDAPL |
User Direct Access Programming Library defines a single set of user-level
APIs for all RDMA-capable Transports. The uDAPL mission is to define a transport-independent
and platform standard set of APIs that exploits RDMA capabilities, such as those
present in IB, VI, and RDDDP WG of IETF. |
| ULP |
Upper layer protocols. APIs for applications to perform IB communications
operations. For instance, MPI-2, IPoIB, SDP, and Sandia Portals. |
| UPC |
Unified Parallel C. An extension of the C programming language designed for
high-performance computing on large-scale parallel machines.The language provides
a uniform programming model for both shared and distributed memory hardware. The
programmer is presented with a single shared, partitioned address space, where
variables may be directly readand written by any processor, but each variable
is physically associated with a single processor. UPC uses a SPMD model of computation
in which the amount of parallelism is fixedat program startup time, typically
with a single thread of execution per processor. |
| URDMA |
Unacknowledged, unreliable RDMA capability. |
| UTR |
University technical representative |
| VAPI |
InfiniBand verbs applications programming interface |
| VP |
Virtual processor. Used in the context of assigning multiple "virtual"
processors to each of physical processors. |
