Title: Building a LowCost Supercomputer
1Building a Low-Cost Supercomputer
NO
- Dr. Tim McGuire
- Sam Houston State University
- ACET 2000
- Austin, TX
2Acknowledgments
- Most treatments of cluster computing (including
this one) are heavily based on the seminal work
of Greg Pfister (IBM Research, Austin,) In Search
of Clusters - The concept of Beowulf clusters originated with
Donald J. Becker and Thomas Sterling at the
Center of Excellence in Space Data and
Information Sciences, NASA Goddard Space Flight
Center
3Introduction
- There are three ways to do anything faster
- Work harder
- "Crunch Time" is familiar to all of us
- Work smarter
- Better to find a way to reduce the work needed
- Get help
- Certainly works, but we all know about committees
...
4In a computer ...
- Working Harder
- Get a faster processor
- Working Smarter
- Use a better algorithm
- Getting Help
- Parallel processing
5Working Harder -- Faster Processors
- The effect of faster processors is astonishing
- The effective speed of the x86 family of
processors has increased nearly 50 per year - RISC architectures have sustained a 60 annual
cumulative growth rate - These trends will likely continue for the
foreseeable future
6Working Smarter -- Better Algorithms
- The increases in speed made possible by better
algorithms dwarf the accomplishments of faster
hardware - Binary search on 1 billion items takes 30
comparisons, versus a maximum of one billion
comparisons using linear search
7Getting Help -- Parallel Processing
- Covert parallel processing
- pipelining, vector processing, etc.
- really equivalent to faster hardware
- Overt parallelism
- Done via software
- "Parallelism is the wave of the future -- and
always will be"
8Early Attempts at Parallelism
- Von Neumann thought it was too hard, and gave us
the "Von Neumann bottleneck" - 60's ILLIAC IV project was the first great
attempt at parallel processing (as well as trying
to advance circuit and software technology.) - Japanese Fifth Generation Project launched
another wave, including the Grand Challenge
problems
9Microprocessor Revolution
- Microprocessors have had a superior
price/performance ratio - "All you have to do is gang a whole bunch of them
together" - The problem is "All you also have to do is
program them to work together" - Programming costs much more than hardware
10Highly Parallel Computing
- Finally, (early 90's) microprocessors became fast
and powerful enough that a practical-sized
aggregation of them seemed the only feasible way
to exceed supercomputer speeds - Even Cray Research (T3D) got into the act
11"Lowly" Parallel Processing
- Mid-to-late 90's -- military downsizing (among
other things) caused funding to dry up - However
- Microprocessors kept getting faster a lot
faster - With overall performance doubling each year, in 4
years what needed 256 processors can be done with
16 instead. - System availability became a mass market issue
- Since computers are so cheap, buy two (or more)
for redundancy in case one fails and use them
both, interconnected by a network
12SMP -- One Form of "Cheap" Parallelism
- Symmetric multiprocessors have been around for
some time and have certain advantages over
clusters - Typically, these have been shared memory systems
-- few communication problems
13The Big Distinction -- Programming
- How you program SMP systems is substantially
different from programming clusters Their
programming models are different - If you explicitly exploit SMP in an application,
it's essentially impossible to efficiently
exploit clusters in the same program
14Why Clusters?
- The Standard Litany
- Why Now ?
- Why Not Now?
15The Standard Litany
- Performance
- Availability
- Price/Performance Ratio
- Incremental Growth
- Scaling
- Scavenging
16Performance
- No matter what form or measure of performance one
is seeking -- throughput, response time,
turnaround time, etc., it is straightforward to
claim that one can get even more of it by using a
bunch of machines at the same time. - Only occasionally does one hear the admission
that a "tad bid" of new programming will be
needed for anything to work correctly.
17Availability
- Having a computer shrivel up into an expensive
paperweight can be a lot less traumatic if it's
not unique, but rather one of a herd. - The work done by the dear departed sibling can be
redistributed among the others (fail-soft
computing)
18Price/Performance Ratio
- Clusters and other forms of computer aggregation
are typically collections of machines that
individually have very good performance for their
price. - The promise is that the aggregate retains the
price/performance of its individual members.
19Incremental Growth
- To the degree that one really does attain greater
performance and availability with a group of
computers, one should be able to enhance both by
merely adding more machines. - Replacing machines should not be necessary.
20Scaling
- "Scalable" is, unfortunately, a buzzword
- What it does deal with is how big a computer
system can usably get. - It is a crucial element in the differentiation
between clusters and symmetric multiprocessors.
21Scavenging
- "Look at all those unused CPU cycles spread
across all the desktops in our network" - Unused cycles are free.
- However, how do you get and manage them? -- this
complicates cluster support very significantly
22The Benefits are Real
- But, how does one take advantage of it?
- The hardware provides the potential.
- The fulfillment lies in the software, and
unfortunately, software isn't riding the
exponential growth curve.
23Why Now?
- Three Trends
- Fat Boxes -- very high performance
microprocessors - Fat Pipes -- standard high-speed communication
- Thick Glue -- standard tools for distributed
computing - One Market Requirement
- High Availability
24Fat Boxes
- Microprocessors have kept, and will keep getting
faster. - Supercomputers in the classic style are extinct
for practical purposes - Mass-market, inexpensive microprocessors have
crawled up the tailpipe of the workstation market
just like workstations crawled up the tailpipe of
minicomputers and mainframes earlier. - There are no more supercomputers, there is only
supercomputing.
25Fat Pipes
- Commodity off the shelf (COTS) networking parts
have achieved communication performance that was
only previously possible with expensive,
proprietary techniques - Standardized communication facilities such as
- ATM - Asynchronous Transmission Mode
- Switched Gigabit Ethernet
- FCS -- Fibre Channel Standard
- Performance of Gigabytes per second are possible.
26Thick Glue
- Standard tools for distributed computing such as
TCP/IP - Intranets, the Internet, and the World Wide Web
- Tool sets for distributed system administration
- PVM (Parallel Virtual Machine) and MPI (Message
Passing Interface)
27Requirement for High Availability
- Nobody has ever wanted computers to break.
- However, never before has high availability
become a significant issue in a mass market
computer arena. - Clusters are uniquely capable of answering the
need of both sides of the spectrum and are much
cheaper than hardware based fault-tolerant
approaches.
28Why Not Now?
- If they're so good, why haven't clusters become
the most common mode of computation? - Lack of "single system image" software
- Limited exploitation
29Lack of Single System Image Software
- Replacing a single large computer with a cluster
means that many systems will have to be managed
rather than one. - Their distributed management tools are tools, not
turnkey systems - 50 of the cost of a computer system is staffing,
rather than hardware, software, or maintenance
30Limited Exploitation
- Only relatively few types of subsystems now
exploit the ability of clusters to provide both
scalable performance and high availability. - This is a direct result of substantial
difficulties that arise in parallel programming. - The problem is not hardware, it's software
31An Exception
- For one kind of parallel system, the software
issues have been addressed to a large degree
The symmetric multiprocessor (SMP) - It of necessity requires a single system image
32Definitions, Distinctions, and Comparisons
- Definition
- Distinction from Parallel Systems
- Distinctions from Distributed Systems
- Comparisons and Contrasts
33Definition
- A cluster is a type of parallel or distributed
system that - consists of a collection of interconnected
stand-alone computers, and - is used as a single, unified computing resource
- We define them as a subparadigm of distributed
(or parallel) systems
34Distinction from Parallel Systems
- A useful analogy
- This is A Dog
- (a single computer)
35A Pack of Dogs
- And this is a pack of dogs (running in parallel)
- (a cluster)
36A Savage Multiheaded Pooch
- or, pardon the abbreviation, "SMP"
- (This pooch is no relation to Kerberos (Cerberus
in Latin) that guards both the gates of Hades and
distributed systems -- He only has three heads.)
37Dog Packs and SMPs are Similar
- Both are more potent than just plain dogs
- They can both bring down larger prey than a plain
single dog. - They eat more and eat faster than a single dog
38Dog Packs and SMPs are Different
- Scaling
- Availability
- System Management
- Software Licensing
39Scaling Differences
- The Savage Multiheaded Pooch can take many bites
at once - What happens when it tries to swallow?
- It needs a larger throat, stomach, intestines,
etc. - Similarly, to scale SMPs, you must beef up the
entire machine - When you add another dog to a dog pack, you add a
whole dog. You don't have to do anything to the
other dogs. - Likewise, clusters
40Availability
- If an SMP breaks a leg
- "that dog won't hunt" no matter how many
heads it has. - If a member of the pack is injured, the rest of
the pack can still bring down prey.
41System Management
- You only have to walk a SMP once.
- It takes a good deal more effort to train a pack
of dogs to behave. - With the SMP, all you have to do is get the heads
to learn basic cooperation (and that should be
built into the operating system.)
42Licensing (Dogs or Software)
- If you get a license for an SMP, you'll probably
only need one license - For an cluster of dogs, you'll need one per dog
43Distinctions from Distributed Systems
- The distinctions of clusters from distributed
systems is not as clear (and a lot of people
confuse the two.) - We'll try. The salient points are
- Internal Anonymity
- Peer Relationship
- Clusters as part of a Distributed System
44Internal Anonymity
- Nodes in a distributed system necessarily retain
their own individual identities - The elements of a cluster are usually viewed from
outside the cluster as anonymous - Internally, they may be differentiated, but
externally the jobs are submitted to the cluster,
not, for example, to cluster node 4
45Peer Relationship
- Distributed systems
- use an underlying communication layer that is
peer-to-peer - at a higher level, they are often organized into
a client-server paradigm - Clusters
- underlying communication is peer-to peer
- organization is also peer-to-peer (with some
minor exceptions)
46Clusters as part of a Distributed System
- Clusters usually exist in the context of a
distributed system - In this case, they are viewed by the distributed
system as a single node - For example, the cluster could server as a
compute engine - It also could serve as, say, a DBMS server in the
client-server paradigm (but that's not the
organization we want to consider in this
presentation)
47Beowulf Clusters
- The Beowulf project was initiated in 1994 under
the sponsorship of the NASA HPCC program to
explore how computing could be made "cheaper
better faster". - They termed this PoPC -- a Pile of PCs
48The "Pile of PCs" Approach
- Very similar to COW (cluster of workstations) and
shares the roots of NOW (network of
workstations,) but emphasizes - COTS (commodity off the shelf) components
- dedicated processors (rather than scavenging
cycles from idle workstations) - a private system area network (enclosed SAN
rather than exposed LAN)
49What Beowulf Adds
- Beowulf adds to the PoPC model by emphasizing
- no custom components
- easy replication from multiple vendors
- scalable I/O
- a freely available software base
- using freely available distributed computing
tools with minimal changes - a collaborative design
50Advantages of the Beowulf Approach
- No single vendor owns the rights to the product
-- not vulnerable to single vendor decisions - Approach permits technology tracking -- using the
best, most recent components at the best price - Allows "just in place" configuration -- permits
flexible and user driven decisions
51Software for Beowulf
- Exploits readily available, usually free software
systems - These systems are as sophisticated, robust, and
efficient as commercial-grade software - Derived from community-wide collaborations in
operating systems, languages, compilers, and
parallel computing libraries
52Operating Systems, etc.
- Two of the operating systems used are
- Linux (Slackware, RedHat, and Debian
distributions are all used) - FreeBSD
- Both have
- commercial distributors and support
- full X Windowing support
- a variety of shells
- a variety of quality compilers
- message passing libraries, such as PVM and MPI
53Beowulf Architecture
- Beowulf clusters have been assembled around every
new generation of commodity CPUs since the first
100 MHz 486DX4 in 1994 - The idea here is to use fast but cheap CPUs
- We also need to interconnect them with a private
network that is fast but cheap - Originally used channel-bonded 10Mbit/sec
Ethernet with multiple Ethernet cards per CPU
because the 100 MHz processors were faster than
the network - When 100Mbit/sec Ethernet cards and switches
became available, channel bonding was discarded
54Alternatives
- Mostly, Intel 80x86 CPUs have been used, but
Beowulf-class clusters have been constructed from
other chips such as the DEC Alpha - Fast Ethernet is most commonly used, but some use
1Gb/sec Ethernet or Myrinet (about 2.5Gb/sec)
where performance is worth the much higher cost
55Topology
- Small systems (-- a single switch
- (If price outweighs performance, a hub may be
used instead of a switch)
Node
Switch
56Connecting to the Outside World
- If the switch is smart (read expensive) it may be
connected directly to the LAN - Most often, however, one node of the cluster has
a second (slower) network card connecting it to
the LAN
57Larger Systems
- Beyond 24 nodes, suitable switches just do not
exist for a single-switch solution - A two level tree with (non-leaf) nodes of 16-way
Ethernet can handle over 200 processors - If locality can be exploited (big problem) there
is no major performance hit - For system wide random communication, the root
node switch can be a severe bottleneck
58Example of a Larger System -- The Hive at NASA
GSFC
- http//newton.gsfc.nasa.gov/thehive
59Beowulf at SHSU -- Bubbawulf
- Bubbawulf consists of 8 nodes
- Pentium 350 with 64MB RAM
- Main node has a 4GB disk
- Other 7 nodes are headless and diskless
- Interconnected through a Cisco 2900 switch (100Mb
full-duplex switched Ethernet network) - The 7 (beowulf2 - beowulf8) mount their file
systems off the main node via NFS
60- Bubbawulf was constructed early in 1999 at a
total cost of about 15,000 and will eventually
be upgraded to the maximum of 24 nodes - Cost now would be about 1/2 of original
- For more on Bubbawulf, see http//beowulf.shsu.edu
61Cheaper Clusters
- The WTAMU Beowulf Project, 1998
- The "Buffalo" CHIP (Cluster of Hierarchical
Integrated Processors) - http//wtfaculty.wtamu.edu/rmashburn
- Total cost
62Hardware
- 16 port Fast Ethernet switch (1000)
- Node 0 (Scavenged -- my 1995 desktop)
- Intel Pentium 90 processor
- 16 MB RAM
- 1.0 GB Hard drive
- 3COM 3c905b 10Mbs Ethernet card (connection to
outside world) - Linksys LNE100TX 100Mbs Ethernet card
- 8x CD-ROM
63Hardware
- Nodes 1-3 (500 each -- mail order)
- Intel Pentium 200 processors
- 32 MB RAM
- 3.2 GB hard drives
- Linksys LNE100TX 100Mbs Ethernet card
- 40x CD-ROM
64Software
- Operating System
- RedHat Linux 6.0
- Message Passing Interface
- LAM MPI version 6.3-3b
65Free Clusters
- The epitome of clusters is the "Stone
Soup-ercomputer" at Oak Ridge National
Laboratories - A group of physicists with no budget built a
Beowulf cluster from cast-off PCs and outdated
network hardware
66The Stone Soupercomputer
67How to Build a "Free" Beowulf
- Gather a bunch of machines that are considered
"too slow" to run Microsoft software - Typically these might be older Pentiums in the 90
- 233 MHz range -- they need not be identical - 16 MB RAM is probably the minimal good amount, 32
MB is better - Hard drives are nice -- diskless stations are
possible, but harder to set up -- 1GB is plenty
big - A CD-ROM simplifies hardware installation
- You'll need at least one monitor and keyboard
68Gather Network Hardware
- Fast Ethernet switches are nice, but not usually
available at low cost - Ethernet hubs are inexpensive, possibly free if
older 10Mbs technology (10baseT) - 10base2 Ethernet is slow, but cheap because it
doesn't require a hub or switch - You will take a big performance hit with the
slower technology
69What Next?
- After you get the hardware set up, start
installing the software - Usually Linux is the OS of choice
- Set up each node as a stand-alone system
- Let them know about each other by assigning IP
addresses (192.168.0.x is a good choice) in
/etc/hosts - Install communication software (MPI or PVM)
70How Does One Program a Beowulf?
- The short answer is Message Passing, a technique
originally developed for distributed computing - The Beowulf architecture means that message
passing is more efficient -- it doesn't have to
compete with other traffic on the net - Other techniques are being explored -- People are
just now looking at Java
71Message Passing Software
- PVM (parallel virtual machine) was first
- Developed at Oak Ridge labs
- Very widely used (free)
- Berkeley NOW (network of workstations) project
72More Recent Message Passing Work
- MPI (Message-passing Interface)
- Standard for message passing libraries
- Defines routines but not implementation
- Very comprehensive
- Version 1 released in 1994 with 120 routines
defined - Version 2 now available
73IEEE Task Force on Cluster Computing
- Aim to foster the use and development of
clusters - Obtained IEEE approval early 1999
- Main home page http//www.dgs.monash.edu.au/rajk
umar/tfcc
74Conclusions
- Cluster computing offers a very attractive cost
effective method of achieving high performance - Promising future
75Quote Gill wrote in 1958(quoting papers back to
1953)
- There is therefore nothing new in the basic
idea of parallel programming, but only its
application to computers. The author cannot
believe that there will be any insuperable
difficulty in extending it to computers. It is
not to be expected that the necessary programming
techniques will be worked out overnight. Much
experimenting remains to be done. After all, the
techniques that are commonly used in programming
today were only won at the cost of considerable
toil several years ago. In fact the advent of
parallel programming may do something to revive
the pioneering spirit in programming which seems
at the present to be degenerating into a rather
dull and routine occupation. - Gill, S. (1958), Parallel Programming, The
Computer Journal (British) Vol. 1, pp. 2-10.