principal architectural developments that support cloud computing are sum- marized in the following sections.
High-Performance Computing
Because of the Internet and high-performance computers, an evolution is occurring in computing. This evolution is the movement from tasks that are computationally intensive to those problems that are data intensive. This evolution characterizes some types of cloud computing applications, which are practical to run because of high-performance computers. These computers play a key role in cloud computing, and some of the major milestones in their development are presented in this section.
The computers known as supercomputers evolved during the 1960s. In 1961, IBM developed the IBM 7030 “Stretch,” which was the fi rst transistor-based supercomputer. It was built for the Los Alamos National Laboratory and was specifi ed at 1.2 MFLOPS (million fl oating-point operations per second.)
High-performance computing and supercomputing cannot be discussed without acknowledging Seymour Cray, who is credited with developing the
c01.indd 11
c01.indd 11 6/24/2010 7:36:22 AM6/24/2010 7:36:22 AM
fi rst “real” supercomputers. While at Control Data Corporation (CDC), Cray developed the 3 MFLOP CDC 6600 in 1964 and the 36 MFLOP CDC 7600 in 1969. These were based on the relatively new silicon transistor technology. Cray left CDC in 1972 to form his own supercomputing company, Cray Research.
CDC continued on the supercomputer path and delivered the 100 MFLOP CDC STAR-100 in 1974. The STAR-100 was a vector processor, meaning it could operate on multiple arrays of data simultaneously.
Supercomputing technology developments accelerated during the next three decades with a variety of products. Detailing every one is beyond the scope of this text, but some of the key machines are summarized in Table 1-2. In the table, Gigafl ops (GFLOPS) represent one billion (109) fl oating point opera- tions per second, Terafl ops (TFLOPS) refer to one trillion (1012) fl oating point operations per second, and Petafl ops (PFLOPS) represent are quadrillion (1015) fl oating point operations per second.
An interesting milestone along the path of supercomputer development was the idea of connecting low-cost, commercially available personal computers in a network cluster to form a high-performance computing system. This idea was formulated in 1993 as the Beowulf computing cluster concept, developed by Thomas Sterling and Donald Becker of NASA. Beowulf uses open-source operating systems such as Solaris or Linux. One of the main characteristics of Beowulf is that all the connected machines appear as a powerful, single resource to the user.
The fi rst prototype in the Beowulf project used 16 Intel DX4 processors con- nected by 10Mbit/second Ethernet. The DX4 processor is an Intel chip with triple clocking. Because the DX4 processor speed was too great for a single Ethernet bus, a “channel-bonded” Ethernet was developed by spreading the communica- tions across two or more Ethernet buses. This approach is no longer necessary with the advent of Gigabit Ethernet. This initial cluster demonstrated the ability of COTS (commercial off the shelf) products to implement high-performance computing systems.
In general, a Beowulf architecture has the following characteristics:
It is designed for parallel computing.
Client nodes are usually diskless, dumb terminals.
Client nodes are connected to a server node through a network, such as Ethernet and Ethernet switches.
It uses Parallel Virtual Machine (PVM) software, which enables multiple networked computers to appear as a single parallel processor.
It uses open-source operating systems such as Linux or Solaris.
It incorporates the Message Passing Interface (MPI) API specifi cation, which enables multiple computers to communicate to form a cluster.
c01.indd 12
c01.indd 12 6/24/2010 7:36:22 AM6/24/2010 7:36:22 AM
Table 1-2: High-Performance Computing Evolution
COMPUTER YEAR PERFORMANCE COMMENTS
Cray-1 1976 250 MFLOPS Used integrated circuits and incorporated around 200,000 logic gates. The fi rst Cray-1 was delivered to Los Alamos National Laboratories.
Cray X-MP 1983 941 MFLOPS A Cray Research parallel vector processor supercomputer, which used multiple processors.
Cray-2 1985 1.9 GFLOPS Successor to Cray X-MP; used an inert liquid cooling system.
Thinking Machines CM-5
1993 60 GFLOPS A Connection Machine model (arrangement of thousands of connected microprocessors) using massively parallel processing and message-passing, distributed memory.
Intel Paragon XP/S
1993 143 GFLOPS Incorporated the Intel i860 RISC microprocessor.
IBS ASCI White 2000 7.226 TFLOPS A computer cluster comprising 512 interconnected IBM commer- cial RS/6000 SP computers.
NEC Earth Simulator
2002 35.86 TFLOPS Comprises 640 supercomputers connected by a high-speed net- work. Each supercomputer con- tains eight vector processors with a peak performance of 8 GFLOPS.
IBM Blue Gene/L
2007 478.2 TFLOPS Based on IBM Power Architecture;
deployed at Lawrence Livermore National Laboratories.
IBM Roadrunner 2008 1.105 PFLOPS The world’s fastest computer, located at the Los Alamos National Laboratory. It incorpo- rates both AMD 64 dual-core Opteron server processors and an IBM Cell accelerator processor that is connected to each Opteron core.
c01.indd 13
c01.indd 13 6/24/2010 7:36:22 AM6/24/2010 7:36:22 AM
Some of the factors that have supported the acceptance and growth of Beowulf- type computers include the following:
Increased demand for affordable, high-performance computing
The availability of open-source software such as Linux
Advances in the development of parallel algorithms
The availability of low-cost, high-speed computer chips used for games, PCs, and entertainment systems
The emergence of fully assembled subsystems for use in clusters
Increased reliability of components and systems
The availability of high-performance computing platforms provides the basis for implementation of cloud computing.