Classification of Distributed Computing Systems

• These can be classified into 4 groups: clusters, peer-to-peer networks,

grids, and clouds.
• A computing cluster consists of interconnected stand-alone computers
which work cooperatively as a single integrated computing resource. The
network of compute nodes are connected by LAN/SAN and are typically
homogeneous with distributed control running Unix/Linux. They are suited to
HPC.
 Peer-to-peer (P2P) Networks
• In a P2P network, every node (peer) acts as both a client and server. Peers
act autonomously to join or leave the network. No central coordination or
central database is needed. No peer machine has a global view of the entire
P2P system. The system is self-organizing with distributed control.

• Unlike the cluster or grid, a P2P network does not use dedicated
interconnection network.

• P2P Networks are classified into different groups:

Distributed File Sharing: content distribution of MP3 music, video, etc. E.g.
Gnutella, Napster, BitTorrent.

Collaboration P2P networks: Skype chatting, instant messaging, gaming

etc.

Distributed P2P computing: specific application computing such as

SETI@home provides 25 Tflops of distributed computing power over 3
million Internet host machines.
 Computational and Data Grids
• Grids are heterogeneous clusters interconnected by high-speed networks.
They have centralized control, are server-oriented with authenticated
security. They are suited to distributed supercomputing. E.g. TeraGrid.

• Like an electric utility power grid, a computing grid offers an infrastructure

that couples computers, software/middleware, people, and sensors
together.

• The grid is constructed across LANs, WANs, or Internet backbones at a

regional, national, or global scale.

• The computers used in a grid include servers, clusters, and

supercomputers. PCs, laptops, and mobile devices can be used to access a
grid system.
 Clouds
• A Cloud is a pool of virtualized computer resources. A cloud can host a
variety of different workloads, including batch-style backend jobs and
interactive and user-facing applications.

• Workloads can be deployed and scaled out quickly through rapid

provisioning of VMs. Virtualization of server resources has enabled cost
effectiveness and allowed cloud systems to leverage low costs to benefit
both users and providers.

• Cloud system should be able to monitor resource usage in real time to

enable rebalancing of allocations when needed.

• Cloud computing applies a virtualized platform with elastic resources on

demand by provisioning hardware, software, and data sets dynamically.
Desktop computing is moved to a service-oriented platform using server
clusters and huge databases at datacenters.
 Advantage of Clouds over Traditional Distributed Systems

• Traditional distributed computing systems provided for on-premise

computing and were owned and operated by autonomous
administrative domains (e.g. a company).

• These traditional systems encountered performance bottlenecks,

constant system maintenance, poor server (and other resource)
utilization, and increasing costs associated with hardware/software
upgrades.

• Cloud computing as an on-demand computing paradigm resolves or

relieves many of these problems.
 Software Environments for Distributed Systems and
Clouds:

Service-Oriented Architecture (SOA) Layered Architecture

• In web services, Java RMI, and
CORBA Stack RMI Stack Web Services Stack
CORBA, an entity is, respectively, a
service, a Java remote object, and a IDL Java WSDL
interface
CORBA object. These build on the
TCP/IP network stack. On top of the CORBA Services RMI UDDI
network stack we have a base Registry
software environment, which would CORBA RMI SOAP Message
be .NET/Apache Axis for web services, Stubs/Skeletons Stubs/Skelet
the JVM for Java, and the ORB ons
network for CORBA. On top of this CDR binary Java native XML Unicode encoding
base environment, a higher level encoding encoding -
environment with features specific to serialization
the distributed computing environment IIOP JRMP HTTP
is built.
RPC or Message Oriented Middleware (Websphere MQ or
JMS)

• Loose coupling and support of ORB JVM .NET/Apache Axis

heterogeneous implementations make TCP/IP/DataLink/Physical
services more attractive than
distributed objects.
 Performance Metrics and Scalability Analysis
• Performance Metrics:
• CPU speed: MHz or GHz, SPEC benchmarks like SPECINT
• Network Bandwidth: Mbps or Gbps
• System throughput: MIPS, TFlops (tera floating-point operations per
second), TPS (transactions per second), IOPS (IO operations per second)
• Other metrics: Response time, network latency, system availability

• Scalability:
• Scalability is the ability of a system to handle growing amount of work in a
capable/efficient manner or its ability to be enlarged to accommodate that
growth.
• For example, it can refer to the capability of a system to increase total
throughput under an increased load when resources (typically hardware)
are added.
 Scalability
• One form of scalability for parallel and distributed systems is:
• Size Scalability
This refers to achieving higher performance or more functionality by
increasing the machine size. Size in this case refers to adding processors,
cache, memory, storage, or I/O channels.
• Scale Horizontally and Vertically
Methods of adding more resources for a particular application fall into two
broad categories:

Scale Horizontally
To scale horizontally (or scale out ) means to add more nodes to a system,
such as adding a new computer to a distributed software application. An
example might be scaling out from one Web server system to three.

The scale-out model has created an increased demand for shared data
storage with very high I/O performance, especially where processing of
large amounts of data is required.
 Scalability
Scale Vertically
To scale vertically (or scale up ) means to add resources to a single node in
a system, typically involving the addition of CPUs or memory to a single
computer.

Tradeoffs
There are tradeoffs between the two models. Larger numbers of computers
means increased management complexity, as well as a more complex
programming model and issues such as throughput and latency between
nodes.
Also, some applications do not lend themselves to a distributed computing
model.
In the past, the price difference between the two models has favored "scale
up" computing for those applications that fit its paradigm, but recent
advances in virtualization technology have blurred that advantage, since
deploying a new virtual system/server over a hypervisor is almost always
less expensive than actually buying and installing a real one.
 Amdahl’s Law
It is typically cheaper to add a new node to a system in order to achieve
improved performance than to perform performance tuning to improve the
capacity that each node can handle. But this approach can have diminishing
returns as indicated by Amdahl’s Law.

Consider the execution of a given program on a uniprocessor workstation

with a total execution time of T minutes. Now, let’s say that the program has
been parallelized or partitioned for parallel execution on a cluster of many
processing nodes.

Assume that a fraction α of the code must be executed sequentially, called

the sequential block. Therefore, (1 - α ) of the code can be compiled for
parallel execution by n processors. The total execution time of program is
calculated by:
α T + (1 - α ) T / n

where the first term is the sequential execution time on a single processor
and the second term is the parallel execution time on n processing nodes.

All system or communication overhead is ignored here. The I/O and

exception handling time is also not included in the speedup analysis.
 Amdahl’s Law
Amdahl’s Law states that the Speedup Factor of using the n-processor
system over the use of a single processor is expressed by
Speedup = S = T / [α T + (1 - α ) T / n]
= 1 / [α + (1 - α ) / n]

The maximum speedup of n is achievable only when α = 0, i.e. the entire

program is parallelizable.

As the cluster becomes sufficiently large, i.e. n  ∞, then S  1 / α, an

upper bound on the speedup S. This upper bound is independent of the
cluster size, n. The sequential bottleneck is the portion of the code that
cannot be parallelized.

Example, α = 0.25 and so (1 – 0.25) = 0.75 then the maximum speedup, S =

4 even if one uses hundreds of processors.

Amdahl’s Law teaches us that we should make the sequential bottleneck as

small as possible. Increasing the cluster size alone may not result in a good
speedup in this case.
 Amdahl’s Law
• Example: suppose 70% of a program can be sped up if parallelized and run
on multiple CPUs instead of one CPU.
• N = 4 processors

S = 1 / [0.3 + (1 – 0.3) / 4] = 2.105

• Doubling the number of processors to N = 8 processors
S = 1 / [0.3 + (1 – 0.3) / 8] = 2.581

Double the processing power has only improved the speedup by roughly

one-fifth. Therefore, throwing in more hardware is not necessarily the

optimal approach.
 System Efficiency
• To execute a fixed workload on n processors, parallel processing may lead
to a system efficiency defined as:

System Efficiency, E = S / n = 1 / [α n + (1 - α ) ]

System efficiency can be rather low if the cluster size is very large.

Example: To execute a program on a cluster with n = 4, α = 0.25 and so

(1 – 0.25) = 0.75,

E = 1 / [0.25 * 4 + 0.75] = 0.57 or 57%

Now if we have 256 nodes (i.e. n = 256)

E = 1 / [0.25 * 256 + 0.75] = 0.015 or 1.5%

This is because only a few processors (4, as in the previous case) are kept
busy, while the majority of the processors (or nodes) are left idling.
 Fault Tolerance and System Availability
• High availability (HA) is desired in all clusters, grids, P2P networks, and
cloud systems. A system is highly available if it has a long Mean Time to
Failure (MTTF) and a short Mean Time to Repair (MTTR).

• System Availability = MTTF / (MTTF + MTTR)

• All hardware, software, and network components may fail. Single points of
failure that bring down the entire system must be avoided when designing
distributed systems.

• Adding hardware redundancy, increasing component reliability, designing

for testability all help to enhance system availability and dependability.

• In general, as a distributed system increases in size, availability decreases

due to a higher chance of failure and a difficulty in isolating failures.