Power Tools 1993 November - Disc 1/Power Tools Plus (Disc 1 of 2)(November 1993)(HP).iso/wkst/50917014/50917014.txt

Text File | 1993-06-28 | 35KB | 656 lines

HP Cluster Computing Program Overview

Background
Cluster computing has become popular as a result of two trends, one
economic and one technical. On the economic side, many Cray and IBM 3090
supercomputers have been leased and the leases are ending. Even
supercomputers that have been paid for cost approximately a million
dollars a year to keep running. As a result, many supercomputing users
are evaluating solutions other than traditional supercomputers to get
their numerically intensive computations done.
   On the technical side, RISC workstations have continually improved in
performance at a very fast rate since their introduction in the mid-
1980s. RISC processor performance leap-frogs nearly every 18 months;
supercomputer performance continues to improve also, but not nearly as
fast. Another important fact is that, while workstations are getting
faster, they are also becoming more affordable. Supercomputers, on the
other hand, are becoming more expensive with time.
   Today, many scalar applications used in supercomputing (such as the
commonly used Monte Carlo simulation) actually run faster on individual
RISC workstations than they do on the fastest supercomputers, since
traditional supercomputers rely on specialized vector processors for
most of their speed. In fact, many users have successfully utilized the
power of multiple workstations to solve their compute-intensive tasks.
HP intends to build on that work and enhance this technology so that
more of our customers can benefit.

Computational Cluster Definition
A computational cluster is a group of two or more workstations networked
together and used as a virtual single computational resource.
Workstation clusters augment the computing power of traditional
mainframes and supercomputers to perform batch or parallel processing.
As shown below, a computational cluster is configured with two or more
headless workstations connected via Ethernet or a higher-speed, lower-
latency networking link such as FDDI. A workstation connected to the
cluster serves as a console to manage the workstations in the cluster.
Software residing on the server workstations manages the task
distribution.

[Figure, page 1: illustration. Captions: Example Workstation Cluster;
 Desktop workstation used as a control console; 8 workstations in a
 1600mm rack]

Cluster Advantages
Workstation clusters are, by their nature, modular and flexible. In the
past, users would actually buy more supercomputing power than they
needed at the time to have room to grow over time. Now, with workstation
clusters, users can simply add more workstations to the cluster as their
needs grow or swap out older systems with more powerful systems. And if
recent history is any indication, the new workstations will not only be
more powerful but also more affordable.
   Many researchers are also looking at clusters as a low-risk entry
into the world of Massively Parallel Processing (MPP). Many believe that
an industry standard will develop with MPP machines and are afraid of
locking into the "wrong" technology. Until a standard emerges, many
researchers are using the workstation cluster as a platform to develop
their parallel programs. This strategy falls in line with HP and
Convex's plans to offer an MPP supercomputer that will be program
compatible with our workstation cluster offerings.

Cluster Scalability
The size of a cluster may vary according to the application and the
processing mode (batch or parallel). The maximum number of workstations
used in a batch processing cluster is unlimited, with current
installations topping 100 workstations. The optimal number of
workstations in a parallel processing cluster ranges from 8 to 16. After
16 workstations, the message passing between machines starts to overcome
the additional useful work being done.

Batch Processing with Clusters
Here a computational cluster acts as a throughput accelerator. This mode
is the most popular way of using computational clusters because there is
no impact on application code. As shown in the following diagram, when a
computational cluster is in batch processing mode, jobs are sent to the
cluster from workstations, X terminals, and other available clients.
Tasks are automatically sent to the cluster via a central queue or a
distributed queue, depending on the batch queueing solution used. Jobs
sent to a queue are large compute-intensive calculations, background
jobs, and other non-interactive tasks which typically require more than
ten minutes processing time. Batch processing is useful for applications
such as CAE, CAD, software development, and scientific analysis. For
example, a software developer uses batch queueing software to run
compiles on a cluster while the developer's own workstation is free to
do more interactive development. The batch software, such as Task Broker
or NQS, performs functions such as task distribution and balancing the
workload to maximize performance.

[Figure, page 2: illustration. Captions: Batch; Lots of Jobs to Lots
 of CPUs; Cluster]

Parallel Processing with Clusters
When a cluster operates in parallel processing mode, as shown in the
following figure, a large job sent to the cluster is divided among
multiple CPUs. Parallel processing is an advanced form of clustering and
requires modifications to the application software to break it up into
relatively independent tasks. The customer uses software tools such as
Linda or PVM to help parse the program into smaller pieces of code or
information. In a parallel processing cluster, interaction between
processors can be slower than in a traditional multiprocessor computer.
Applications must have a high ratio of compute-intensive calculations to
I/O communications to perform well in a cluster computing environment.
This type of application is called a coarse-grained application.
Molecular dynamics, raytracing used in visualization, and seismic data
analysis are examples of coarse-grained applications.

[Figure, page 3: illustration. Captions: Parallel; One Job to Lots of
 CPUs; Cluster]

The Convex/HP Partnership
HP has extended its partnership with Convex Computer to include Cluster
Computing. Convex is a leader in the supercomputer market and provides a
high level of expertise in parallel processing. Convex is a member of
PRO (Precision RISC Organization) and will develop MPP (Massively
Parallel Processing) machines using the PA-RISC chip. Convex has ported
three cluster computing enabling products to the Series 700 workstation.
The products are ConvexNQS+, ConvexPVM, and ConvexMLIB. HP and Convex
are pooling knowledge and resources to develop a suite of scalable
computing products to meet the needs of the compute-intensive market.
The result of this relationship is a four-tiered product offering for
computational clusters.

[Figure, page 3: illustration. Captions: Cluster Solutions; Desktop
 Cluster Existing Workstations on a LAN; Dedicated Cluster of
 Workstation Servers; Pre-assembled Cluster; Meta Series Convex
 C Series Super Computer with Series 700 Workstations]

   1. Desktop Cluster: Here, with the addition of the right batch or
parallel processing software, existing desktop workstations are utilized
as a single computational resource in off-hours to do batch or parallel
processing in addition to the interactive design work they were
purchased for.
   2. Dedicated Cluster: Many people have already set up dedicated
clusters of segregated workstations in addition to those on the desktop
to off-load the very computationally intensive parts of their work to
free up the interactivity of the desktop client. Typically several high-
performance workstation servers are set up for this task.
   3. Pre-assembled Cluster: Customers now have the option to purchase a
cluster already assembled in a rack. Configurations of up to eight
Series 700 workstations can be mounted into a rack with the system
already pre-configured with the selected network (Ethernet or FDDI)
installed.
   3. Meta Series: Convex offers a "Metacomputer" containing a C series
supercomputer and a cluster of 2-8 HP workstations. The Meta Series
integrates the vector processing and large data handling capabilities of
a supercomputer with the scalar processing of workstations and is
appropriate for customers who require both processing methods. This
product is sold by Convex.

Software for Clusters

Software for Batch Processing
There are several software programs available on the Series 700 to
manage a cluster network in batch processing mode: Task Broker, NQS,
ConvexNQS+, and Load Balancer.

Task Broker

Hewlett-Packard
Task Broker is an HP product available on all Series 700 workstations,
as well as the Series 800 servers. Task Broker is also available from a
third-party vendor, SAIC, to run on Sun SPARCstations. Task Broker
maximizes the performance of the workstation cluster by efficiently
distributing tasks to the best server available.
   The process by which a task is distributed to the cluster is as
follows:
   1. Once installed, Task Broker creates "daemons," which run in the
background waiting for service requests.
   2. Tasks (service requests) are submitted from the user's machine.
   3. The "daemon" on the user's machine queries the "daemons" on other
workstations in the cluster about their ability to perform the task. The
criteria for selecting a server include: the type of application, the
time of day, the number of services currently using the server, the CPU
load, and the available disk space.
   4. The daemon on the user's workstation selects the workstation that
is both available and best suited to perform the task.
   5. If all servers are busy, the task waits in a queue until a server
is free.
   6. Files needed to complete the task may be transferred to the server
by Task Broker or accessed remotely via NFS.
   7. The task is completed on the selected workstation, and the results
are sent back to the originating workstation.
   Task Broker resides on each client and server workstation in the
cluster. The version of Task Broker must be appropriate for that model
server. There is no centralized server which controls the brokering and
queueing of tasks. Important features in Task Broker include:
   o Application code does not have to be recompiled to be run by Task
Broker.
   o Task Broker automatically sends results back to a user-specified
file on the user's computer. Mail messages can be sent to notify the
user the job is complete.
   o Servers and clients are added to a network equipped with Task
Broker simply by hooking the machine to a LAN, updating the central Task
Broker configuration file to include the new client/server, and starting
the Task Broker daemon.
   o Task Broker provides an accounting of services used on a given
server.
   o The number of applications running on a computer at any one time
can be limited. This prevents degradation of server performance.
   o Users can control the use of their computer as a server. For
example, users can specify that their workstations be accessed only at
night.

Task Broker Benefits
In situations where jobs need to be executed on remote computers, Task
Broker offers more flexibility and power than existing network
utilities, such as those for remote copying of files (ftp) or remote
shell execution (remsh). Moreover, Task Broker enables you to define
what services are to be provided on which computers, how and when each
service will be provided, as well as who can access these services. Task
Broker has the following user benefits:
   More efficient access to compute resources (i.e., load balancing) -
Task Broker assigns jobs to the server node, whether a specialized
server platform or unused workstation - whichever is most appropriate at
the time. For example, a high-end compute server is obviously better for
solving compute-intensive simulations such as finite element analysis.
However, if this server is overloaded and a workstation that could do an
adequate job is left idle, the overall productivity of the computer
environment suffers.
   Improved productivity and products - By allowing multiple compute
jobs to run in parallel and/or on faster systems, performance is
dramatically improved. Users may use the time saved to run more jobs in
a given time period in order to improve product design or shorten
development time.
   Greater flexibility and ease of accessing multiple servers - The fact
that all users get access to each other's idle cycles makes for a
winning proposition for the entire work group.
   Access to heterogeneous computers - You can build a cooperative
computing environment. For example, a task submitted from an HP 9000
Series 700 workstation may be serviced by a Series 300 or 400
workstation, or vice versa.

Features of the New Release
   o A graphic user interface (GUI) has been added, greatly improving
the product's ease-of-use. The GUI provides a visual interface to most
of the Task Broker command set and configuration information,
simplifying Task Broker administration. In addition, task status
monitoring and control is provided for the end user.
   o Centralized configuration management has been added to the new
release. This feature allows the entire Task Broker installation to be
initialized using a single group configuration, and to be administered
from any single machine site.
   o An integrated forms-based configuration editor provides for easier
and more robust administration of Task Broker information. The
configuration syntax is simpler and checking is done during the editing
session.
   o An online, context sensitive help sub-system, utilizing HP's CACHE
CREEK product, contributes to Task Broker's overall ease-of-use by
providing usage information when it's needed.0

NQS and NQS/Exec

Sterling Software
NQS (Network Queueing Software) is a UNIX-based product available from
Sterling Software that facilitates batch processing on a computational
cluster. NQS is a batch queueing program that allows the user to submit
a job to the cluster via a system of queues. The system administrator
creates the queues according to the system demands and computing
resources available. NQS/Exec adds automatic load leveling to the NQS
product. With NQS/Exec, when the user submits a job to the queue, the
software automatically selects the best processor available to complete
the task. NQS will not move the data over to the workstation
automatically unless the file system has been previously mounted via
NFS. Additional features of NQS and NQS/Exec include:
   o POSIX batch-queueing standard
   o Interoperability with multiple platforms
   o Ability to submit a job to a selected queue and track the progress
of the task
   o The ability to monitor server use

ConvexNQS+

Convex Computer Corporation
NQS+ is a batch queueing product based on NQS and supported by Convex.
ConvexNQS+ offers the standard features of NQS:
   o Job routing
   o Job accounting
   o Access control
   o Queue limits
   o Standard interconnects and protocols
   ConvexNQS+ allows users on a network to submit jobs into a batch
queue on the Cluster. ConvexNQS+'s load balancing feature can
automatically route jobs to the most appropriate processor in the
cluster based on job priority, average system load, processor speed, and
the length of queue. ConvexNQS+'s file importing feature allows users to
execute jobs without having to manually copy files to the system
processing the job. ConvexNQS+ remote mounts the user's working
directory onto the executing system, makes sure the data is there, and
un-mounts after job completion. Therefore, ConvexNQS+ not only
automatically load balances, but also can automatically export user's
data file to any node in a cluster. ConvexNQS+ is compatible with NQS
and allows users to submit jobs to and from other systems that support
NQS. ConvexNQS+ provides diagnostic messages that keep users and
administrators informed about job and queue status.
   ConvexNQS+ also offers the following additional functions:
   o direct remote submission and removal of jobs
   o load balancing
   o file importing
   o fair share scheduling
   o demand or "pull" queues (more efficient than "push" queues)
   ConvexNQS+ is available exclusively on HP Apollo Series 700
workstations and Convex C-series supercomputers.

Load Balancer

Freedman Sharp and Associates Inc
Load Balancer from Freedman Sharp and Associates is a batch queueing
product which operates in heterogeneous networked environments. Load
Balancer is available on both the Series 700 and 800, as well as on Sun,
DEC, IBM, and SGI workstations. Load Balancer offers a central
configuration file for simpler system administration and maintenance. In
addition, Load Balancer allows the user to customize the configuration
down to the swap space needed by an application on a specific
workstation. Other key features in Load Balancer include:
   o automatically queues and distributes jobs
   o 256 job priority levels
   o queues, pauses, restarts, or defers batch jobs
   o understands fixed and floating licenses, access control and node
availability schedules
   o tracks server usage and performance

Software for Parallel Processing
There are several software programs available that enable parallel
processing on a distributed network. The software programs available on
the HP Apollo Series 700 workstations are: ConvexPVM, Express, Forge 90,
ISIS, Network Linda, PVM, and Tuplex,.

Convex PVM
Convex Computer Company
Based on PVM, ConvexPVM enables parallel execution of a single
application across multiple processors in the cluster. With ConvexPVM,
users can explicitly embed library calls within their C or FORTRAN
application that enable parallel execution. At runtime, ConvexPVM
coordinates and synchronizes initialization, process spawning, and
message passing between the processors in the cluster as they execute
the submitted job. In addition, ConvexPVM is also tuned for the CONVEX
and PA-RISC architectures and additional routines have been added that
provide a better user interface.

ConvexMLIB
Convex Computer Company
Based on Convex's famous VEClib package that runs on their C Series
supercomputers, ConvexMLIB provides a comprehensive set of subroutines
that can tackle the most demanding tasks found in numerically intensive
computing. For users who want to migrate code from a Cray supercomputer,
ConvexMlib includes a Cray SCILIB compatible library. Sections of
ConvexMLIB routines have been assembly coded for maximum performance.
Future versions of ConvexMLIB will be parallelized with ConvexPVM so
parallel processing across workstations in a cluster can be
transparently invoked with a simple library call.

Express
ParaSoft Corporation
Express is a programming environment marketed by Parasoft. Express
includes a five-phase procedure in the conversion of a sequential
application to parallel application. The phases are:
   1. Project evaluation
   2. Code development
   3. Debugging and testing
   4. Performance evaluation
   5. Maintenance and upgrades
   The code-development phase of Express contains a tool called ASPAR,
which aids the user in converting sequential applications code to
parallel code. Express is available for applications written in C or
Fortran.

FORGE 90

Applied Parallel Research, Inc.
FORGE 90, from Applied Parallel Research, Inc., is an interactive
analysis environment for the exploration and parallelization of
industrial strength Fortran programs. FORGE 90 is a series of layered
products composed of the following:
   o Baseline FORGE 90 is a an interprocedural Fortran program
browser/analyser, providing tracing and in-context search functions
across subprograms and through calls and COMMON. FORGE 90 locates
undefined and useless variables, scrutinizes COMMON block usages,
displays the relationships between routines, blocks and statements
(including data dependencies), and identifies the CPU intensive parts of
a program down to the DO loop level. This tool is instrumental in
providing information necessary for parallelization and optimization.
   o Parallelizer for Distributed Memory Systems (DMP) is an interactive
tool used with Baseline FORGE 90 to parallelize Fortran applications.
DMP works with the information gathered by baseline FORGE 90 to assist
the user in parallelizing high level loops (to optimize efficiency). The
result is a Fortran 77 SPMD (Single Program, Multiple Data) program with
parallelized DO loops and partitioned data arrays spread across
processors, and with communications and synchronization calls to APR's
parallel runtime library PVM, Express or Linda interface. DMP also
contains a parallel runtime profiler, which identifies performance
bottlenecks and communications overloads, and predicts performance on
scalable systems based on single node performance.
   o xHPF and xHPF77 are batch command tools which convert programs with
High Performance Fortran (HPF) Subset Data Decomposition Derivatives and
Fortran 90 Array syntax or Fortran 77 DO loops into SPMD parallelized DO
loops with calls to APR's runtime library.
   FORGE 90 is available on most Unix workstation platforms.

ISIS

ISIS Distributed Systems, Inc.
ISIS is a product line of distributed computing tools developed by ISIS
Distributed Systems, Inc. ISIS Distributed Toolkit provides a software
programming environment to develop distributed applications in C, C++,
Fortran, Common LISP and Ada. ISIS offers software fault tolerance to
the distributed computing environment. Additional products which can be
layered on to the Toolkit are--
   o ISIS Distributed Resource Manager: This product is used to create a
fault-tolerant "supercomputer" out of a network of workstations. The
ISIS Distributed Resource Manager actually enables load-balanced batch
processing in a distributed computing environment.
   o ISIS Distributed News: Designed for banking and brokerage
applications, this product is a fault-tolerant message publication-
subscription system that serves as a communications backplane for
applications requiring subject-based broadcasting.
   o ISIS Reliable Network File System: This product is a fault-tolerant
network file system based on NFS protocol. Completely compatible with
NFS, the ISIS RFNS product requires no changes to the applications
software, or the NFS servers, and no additional hardware.
   o ISIS Distributed Sensors: This product is a system administrative
tool designed to monitor server performance and availability.

Network Linda

Scientific Computing Associates
Network Linda is a parallel processing language sold by Scientific
Computing Associates (SCA). Network Linda contains six commands that are
inserted into applications programs (available for C, C++, and Fortran)
to run independent tasks simultaneously. SCA markets Network Linda to
customers with high-level computing needs who have developed their own
applications programs. Examples of these customers in the commercial
market include the oil, gas, and finance industries.
   Network Linda operates on a "Master/Worker" paradigm. The "Master"
breaks the application into tasks via Linda commands embedded in the
source code. When the program runs, the "worker" grabs a task, completes
it, returns it to the master, and grabs another task. Tasks are
contained in a unit called a "Tuple," and several tuples together are
called a "Tuple Space". Tuples may also contain data or results. Network
Linda coordinates the interaction between Tuples so that the parallel
operation is invisible to the end user. The only visible difference is
that the application runs several times faster.

PVM/HeNCE

Oak Ridge National Laboratory
PVM (Parallel Virtual Machine) is a software environment coordinating
parallel processing in a heterogeneous network. Developed by the Oak
Ridge National Laboratory in Tennessee, PVM is public domain software
available by sending electronic mail to "[e-mail]" with the
message "send index from pvm". PVM achieves parallel processing via a
message sending and receiving paradigm. PVM is well suited for both
applications with interrelated subtasks and for traditional parallel
applications (those without a great deal of interaction). The Oak Ridge
Lab developed HeNCE to serve as a graphical interface tool and
methodology to use PVM. PVM/HeNCE contain the tools to configure a
heterogeneous network, manage multiple object modules in each component
of the application, execute the application in parallel mode, debug, and
monitor performance.

Tuplex

Torque Systems
Tuplex is a Linda based parallel programming environment developed
specifically for commercial applications. Torque works with application
developers to parallelize third party applications for business imaging,
spreadsheet/financial analysis, and other highly parallelizable
applications. Like Network Linda, Tuplex uses the "Master/slave"
paradigm to achieve the distribution of tasks across a cluster.

System Administration and Management Software
The management and administration of a cluster are key issues in a
cluster environment. HP offers system administration and management
software to aid in critical network administration practices, such as
back-up, recovery, and performance management in a cluster. The tools
from HP that provide these services are Network Node Manager, PerfView,
GlancePlus, RemoteWatch, OmniBack, and OmniBack/Turbo.

Network Node Manager and PerfView
Network Node Manager, available as part of PerfView, dynamically
generates and maintains a map of the network. It can be used to collect,
analyze and graph data about network layout and usage. PerfView is a
system-management software tool that collects data on and analyzes
network performance. Using a color graphical map, PerfView displays the
CPU, disk, and LAN utilization in the cluster. PerfView contains
sophisticated alarm-filtering technology to identify potential problems
in the cluster.

HP GlancePlus
GlancePlus is used for online monitoring of individual systems. Each
system's activity within the cluster can be displayed within a window.
It can characterize system performance and resolve problems. GlancePlus
monitors CPU, memory, and I/O usage by process and by user. GlancePlus
can track historical trends and it can help balance disk use and tune
file systems.

OmniBack and OmniBack/Turbo
OmniBack is an automated global network backup management solution that
provides sophisticated scheduling and journaling facilities to make
centralized backup and recovery on a cluster a simpler procedure.
OmniBack/Turbo is a high-speed version of OmniBack used to provide
backup services for users with large databases (using local raw disk).
OmniBack/Turbo is available on both the HP Apollo Series 700
workstations and the Series 800 business servers, and is especially
valuable in clusters with an 800 as a high-availability file server.

Example Pricing
Here is an example cluster configuration based on eight Series 700 Model
735 headless workstations.
   Cluster Processing Module x 8
            735 Server
            32MB RAM, 1.0GB Disk
   Cluster Control System
            715/50 Color Workstation, EISA
            32MB RAM, 1.0 GB Disk
   Rack Cabinet
            1600mm rack for 8 systems
   System Integration
   Total:   $315,940 (US List)

Software Pricing
Here is US List pricing for the software. Contact each vendor directly
for detailed pricing.

                                             Unit Price
Company                    Product           (Per node)  10 Nodes
------------------------------------------------------------------------
 Convex                    ConvexNQS+        $1,250      $4,750*
                           ConvexPVM         $2,600      $9,500*
                           ConvexNSQ+and
                           ConvexPVM         $3,375      $12,825*
                           ConvexMLIB        $4,995      $25,000*
                                                         * 8 node price

 Sterling                  NQS 2.2           $495        $4,700
                           NQS/Exec 2.2      $795        $7,550

 HP                        TaskBroker         ---        $5,000

 Freedman Sharp & Assoc.   Load Balancer     $299        $1,500
                                                         (Site License)

 Scientific Computing      Network Linda     $1,000      $4,995
 Assoc

 Isis Distributed          Isis Distributed  $500        $5,000
  Systems                  Toolkit           $2,500      $25,000

 Oak Ridge National Lab    PVM               $0          $0
                           HeNCE             $0          $0

 Parasoft                  Express                       Contact
                                                         Parasoft

 Torque                    Tuplex                        Contact Torque

 Applied Parallel Research FORGE 90          $2,405      $22,295

        Contact Summary of Third-party Cluster Software Providers

Product     Contact        Company                    Phone
------------------------------------------------------------------------
 Linda      Leigh Cagan    Scientific Computing       203-777-7442
                           Associates                 203-776-4074 (FAX)
                           One Century Tower
                           265 Church Street
                           New Haven, CT 06510

 NQS        Thomasine      Sterling Software          415-964-9900
            Bailey         1121 San Antonio Road      415-969-3821 (FAX)
                           Palo Alto, CA 94303

 ConvexNQS+ Blair Baker    Convex Computer Corp.      214- 497-4000
 ConvexPVM                 3000 Waterview Parkway     214-497-4848 (FAX)
 ConvexMLIB                P.O. Box 833851
                           Richardson, TX 75083

 Express    Bill Rodart    Parasoft                   818-792-9941
                           2500 E. Foothill Blvd      818-792-0819 (FAX)
                           Suite 205
                           Pasadena, CA 91107

 Isis       Rick Moran     Isis Distributed Systems   212-979-7729
                           Suite 200                  607-272-6327
                           111 Cayuga Street          607-277-2611 (FAX)
                           Ithaca, NY 14850

 Load       Dan Freedman   Freedman Sharp and Assoc.  403-251-2729
 Balancer                  204 Woodglen Place S.W.    403-281-0204 (FAX)
                           Calgary, Alberta T2W 4G3

 PVM        netlib         Oak Ridge National
            @ornl.gov      Laboratory
            (e-mail)       Oak Ridge, TN

 Tuplex     Scott Rafer    Torque Systems, Inc        415-321-1200
                           825 Emerson Street         415-321-1298 (FAX)
                           Palo Alto, Ca 94301

 FORGE 90   Bob Enk        Applied Parallel           301-718-3733
                           Research                   301-718-3734 (FAX)
                           5500 Lambeth Road
                           Bethesda, MD 20814         or
            John Levesque  550 Main Street, Suite 1   916-621-1600
                           Placerville, CA 95667      916-621-0593 (FAX)

Cluster Lab
We have built a lab in Fort Collins, Colorado, that houses a cluster of
735 workstations with all the third-party enabling software. The lab
will be able to demonstrate the capabilities of a workstation cluster
and will also be available for benchmarking application code. HP will
use this lab to characterize and report on the performance of various
applications on workstation clusters using different third-party
programs. We will also experiment to see how different workstation
configurations and topologies affect performance. Convex has a similar
lab in Richardson, Texas.

Future Directions
HP intends to build on this first cluster offering by investing in
technologies that will improve system performance and make clusters
easier to use, administer, and manage.
   A new release of Task Broker will be available in mid 1993. The new
release will have a graphical user interface, a central configuration
management file, and improved task status and control tools.
   Areas of investigation for future product release include higher
bandwidth, lower latency connections between machines in a cluster. In
addition to a faster FDDI implementation on the 7100-based workstations,
products under development include HiPPI (High Performance Parallel
Interface), Fibre Channel and ATM (Asynchronous Transfer Mode). Due out
by the end of 1993, HiPPI provides an 800 Mb/second bandwidth. Fibre
Channel is designed to communicate between clusters and mainframes at
speeds ranging from 266 to 1062Mb/second. Fibre Channel is expected to
be available in the 1993-94 time frame. ATM offers very low latency and
speeds ranging from 45Mb/sec to 155Mb/sec, integrates voice and data,
and is well suited for applications containing multimedia. ATM is
expected to be available in the 1994 time frame.
   In addition, we will be focusing HP's already ongoing development in
network computing and network management technologies on cluster
computing to make clusters easier to administer and more resilient
during operation. Process migration, checkpoint restart, and single
system image are two important features under development.
   Convex will play an active role in future developments in the cluster
program. Convex is planning to build MPP (Massively Parallel Processing)
machines using PA-RISC CPUs. The Convex MPP machines will be compatible
with existing HP Series 700 clusters using Convex software. The
advantage to the HP-Convex cluster customer is that the MPP machines
will be able to provide more power without requiring the customer to
change software. Convex is planning to release its first MPP machines in
1994.

UNIX is a registered trademark of UNIX System Laboratories Inc. in the
U.S.A. and other countries.
   The information contained in this document is subject to change
without notice.

Copyright (c) Hewlett-Packard Co., 1993
Printed in U.S.A. 4/93
5091-7014E