Objectives of these slides:

1
2. Server Architecture

• Objectives of these slides
• examine suitable hardware architectures for
  servers.
• Background Reading
• Chapter 4 on Berson
• Operating Systems , Computer Architectures

2
Overview

• 1. Server Functions
• 2. Server Requirements
• 3. Hardware Choices
• 4. Why Multiprocessor Systems?
• 5. Server OS Requirements

3
1. Server Functions

• Resource sharing and management
• file sharing
• printer sharing
• Database manipulation
• Communication services

4
2. Server Requirements

• Multi-user support
• Scaleability
• vertical and horizontal
• Performance, throughput
• Storage capacity
• Availability
• robustness, on-line administration
• Complex data types
• Networking and communication

5
2.1. A Performance Measure

• Performance 1 / ( CT PL CI)
• CT (cycle time) ? inverse of clock rate
• faster clock ? reduced cycle time
• PL (path length) no. of instructions required
  to execute one command
• CI (cycles/instructions) no. of cycles
  necessary to execute one instruction

6
3. Hardware Choices

• Mainframes were traditional used as servers, but
  modern microcomputers have much better
  price/performance measures
• SPARC 1,500 transaction/sec (TPS)
• IBM 150,000 TPS
• But, SPARCs are slower, and come with less
  well-developed transaction processing software.

continued
7

• Three popular server architectures
• RISC (Reduced Instruction Set Computer)
• SMP (Symmetric MultiProcessing)
• MPP (Massively Parallel Processing)

8
4. Why Multiprocessor Systems?

• Specialised processors
• FPU, graphics
• Parallel execution of large databases
• Concurrent processing of users
• Higher transaction rates
• Complex applications
• virtual reality, weather forecasting

9
4.1. Classifying the Architectures

• Interconnection structure between processors
• bus, star, ring, etc.
• Component interdependencies
• Degree of coupling between processors and memory
• shared memory, distributed memory

continued
10

• Architectures can be classified by the way that
  instructions (code) and data are
  distributed.Two main designs
• MIMD
• SIMD

11
4.2. Shared Memory Architecture
Global(shared)memory
Global Bus

• bus must be wide (64 or 128 bits)
• could be in one machine or many

12
Features

• Large memory
• 4-30 processors
• If all the processors have similar capabilities,
  then called a symmmetric multiprocessor (SMP)
  architecture.
• Seemless distribution (to the users)
• Limited scaleable performance
• system bus, memory, cache size, cache coherence

continued
13

• Fairly easy to program since memory model is
  similar to a uniprocessors
• techniques include UNIX forks, threads, shared
  memory
• Many application areas involving parallel, mostly
  independent, tasks
• e.g. ATM funds withdrawal
• Implementations
• Sequent Symmetry, Sun SparcCenter

14
Programming Shared Memory

• Can use conventional languages with simple
  concurrency and locking of shared data

Global (shared)memory

• Where to put processes?
• Many-to-many communication

g
1.0
g g1
g g2
processes
15
4.3. Distributed Memory Architecture

• network could be inside one machine or many

16
Features

• 100s of processors massively parallel
• New programming paradigms are needed
• new OSes, new compilers, new programming
  languages
• Programming languages are often based on
  processes (concurrent objects) and message
  passing.

17
Programming with Distributed Memory

• Harder to program since must explain how
  processes communicate.

• Poor UNIX tools
• Where to put processes?

processes
18
4.4. MIMD

• Multiple Instruction, Multiple Data
• Multiple processors simultaneously execute
  different instructions on different data.
• Best for large-grained parallel applications
• e.g. concurrent objects communicating
• Can be employed by shared-memory and distributed
  architectures.

19
4.5. SIMD

• Single Instruction, Multiple Data
• Multiple processors simultaneously execute the
  same instructions on different data.
• Best for fine-grained parallel applications
• e.g. parallel pixel modification in a graphic,
  many mathematical applications
• Can be employed by shared-memory and distributed
  architectures.

20
5. Server OS Requirements

• Support for multiple concurrent users
• Preemptive multitasking
• a (higher priority) task/process can interupt
  another
• Light-weight concurrency
• threads, (forks)
• Memory protection
• Scaleability
• load balancing

continued
21

• Security
• Reliability/availability
• Two examples
• UNIX, Windows/NT

22
3. Language Constructs for Parallelism

• Objectives of these slides
• give an overview of the main ways of adding
  parallelism to programming languages

23
Overview

• 1. Relevance to Client/Server
• 2. Co-routines
• 3. Parallel Statements
• 4. Processes
• 5. Linda Tuple Spaces
• 6. More Details

24
1. Relevance to Client/Server

• Most server architectures are multiprocessor
• e.g. shared-memory, distributed architectures
• Most server applications require parallelism
• e.g. concurrent accesses to databases
• Client applications can also benefit from
  parallel coding
• e.g. concurrent access to different Web search
  engines

25
2. Co-routines

• A routine is a function, procedure, method.
• Two co-routines are executed in an interleaved
  fashion
• part of the first, part of the second, part of
  the first, etc.
• A co-routine may block (suspend) and resume
  another co-routine.

26
Example

• void odd() int i for (i1 i lt 1000
  ii2) printf(d, ii) resume
  even()

void even() int i for (i2 i lt 1000
ii2) printf(d, ii) resume odd()

27
Features

• Very hard to write co-routines correctly since
  the programmer must remember to give all of them
  a chance to execute.
• Easy to starve a co-routine
• it never gets resumed
• Deadlocked programs are ones where no
  co-routines can proceed
• perhaps because they are all suspended

28
3. Parallel Statements

• parbegin statement1 statement2
  statement3parendparfor(i0 ilt10 i)
  statement using i
• Other constructs are possible
• parallel if

29
Examples

• Add two arrays (vectors) in parallel
• int a10, b10, c10 parfor(i0 i
  lt 10 i) ci ai bi

continued
30

• Parallel matrix multiplication
• float x1010, y1010, z1010
  parfor(i0 ilt10 i) parfor(j0 jlt10 j)
  zij 0.0 for(k0 k lt10 k)
  zij zij xik
  ykj

31
Features

• Very fine-grained parallelism is typical
• can adjust it to some degree
• Most common as a way of parallelising numerical
  applications (e.g. in FORTRAN, C).
• Automatic parallelisation is a research aim
• e.g. compiler automatically changes for to parfor

continued
32

• Speed-ups depend on code, data and the underlying
  architecture
• creation of processes may take longer than the
  execution of the corresponding sequential code
• data dependencies may force a sequential order
• x 2y x 3
• no point having 100 parallel processes if there
  are only 2 actual processors

33
4. Processes

• A process is a separately executing piece of
  code. Each process can have its own data.
• The main issues are how processes communicate and
  how they share resources.
• Threads are light-weight versions of processes
• they store less OS information
• they have simpler communication facilities

34
Process Example

• void main() fork compile(file1.c,file1.o)
  fork compile(file2.c, file2.o) join
  link(file1.o, file2.o)void compile(char
  fnm, char ofnm) . . .

35
Comunication Synchronisation

• Example a parallel compiler

front_end()
source code
assembly code
object code
assembler()
36
4.1. Shared Variables

• int count 0 / shared data /Line
  bufferMAX / shared data /process void
  front_end() int write_pos 0 Line ln
  while (true) / generate a line of assembly
  / while (count MAX) / busy wait /
  bufferwrite_pos ln write_pos
  (write_pos1) MAX count

continued
37

• process void assembler() int read_pos 0
  Line ln while (true) while (count
  0) / busy wait / ln
  bufferread_pos read_pos (read_pos1)
  MAX count-- / process line of assembly
  code /

38
Issues

• Busy wait versus polling
• wait until count lt MAX
• Race conditions
• seen in updates to count
• solution lock variables, semaphores, monitors,
  etc.

39
4.2. Message Passing

• process void front_end() Message msg Line
  ln while (true) / generate a line of
  assembly code / msg ln send msg to
  assembler()

process void assembler() Message msg Line
ln while (true) receive msg from
front_end() ln msg / process line of
assembly code /
40
Issues

• Should the sender block (suspend) until a message
  arrives at its destination?
• how does the sender know it has arrived?
• How can the sender receive a reply?
• how is a reply associated with the original
  message?
• should a sender wait for a reply?
• Message ordering
• Testing received messages
• how are messages stored while being tested?
• what happens to a message that is rejected by a
  receiver?
• RPCs are a form of message passing where the
  sender waits for an answer (the result of the
  function call).

41
5. Linda Tuple Space

• Linda is the name for a set of operations that
  can be added to existing languages.
• e.g. C/Linda, FORTRAN/Linda
• The operations allow the creation/manipulation of
  a tuple space by processes.

continued
42

• A tuple space is a form of (virtual) shared
  memory consisting of tuples (like structs,
  records) that are accessed by matching on their
  contents.
• Three atomic operations
• out adds a tuple to the space
• read reads in an existing tuple
• in reads in and deletes the tuple from the
  space
• read and in may block the process using them

43
Example
process 1

• int x read(matrx,2,3, x)

process 2
out(matrx,2,2, 3.4)out(matrx,2,3,
3.1)
44
6. More Details

• Programming Language EssentialsH.E. Bal D.
  Grune Addison-Wesley, 1994
• A more technical and detailed discussion
• Models and Languages for Parallel
  ComputationD.B. Skillicorn, D. TaliaACM
  Computing SurveysVol. 30, No. 2, June 1998