Transcript IP address - UMM Directory

COMPUTER
COMMUNICATION
Sunggu Lee
EE Dept., POSTECH
Apr. 22, 2009
Computer Communication Example

Send picture image and message to friend
sender
receiver
Microsoft Outlook
Hello!
Netscape Messenger
Hello!
Hello!
system software
system software
Communication
Channel
Packetization of Data

For transmission of a stream of data bits (message),
the message is typically partitioned into “packets”
A
packet consists of (at the very least)
 Packet
header (destination, routing info, etc.)
 Data payload (the bits of the message)
 Check bits (redundant bits used to check for errors in the
received packet)
Communication Protocols

For successful transmission/receipt of a packet, the
transmitter and receiver must agree on a
“communication protocol”

Set of rules on how the packet is interpreted

How to sample the bits of the packet




Signaling method
Synchronization of the transmitter/receiver
How to determine which parts of the packet are the packet header
(destination info, etc.), data payload, check bits, etc.
How to interpret the bits of the data payload

Integer, floating-point, character string, JPEG picture, etc.
Computer Communication Models and
Communication Protocol Suites

Most commonly used reference base communication
model is the Open Systems Interconnection (OSI)
model


Standardized by the International Organization for
Standardization (ISO)
Most common implementation of the OSI model is a
set of protocols referred to as the TCP/IP protocol
suite (or stack)
TCP = Transmission Control Protocol
 IP = Internet Protocol

Communication Protocols
OSI Model
L7
Application
Application
L6
Presentation
Presentation
L5
Session
Session
L4
L3
L1
SMTP FTP
Data
DataLink
Link
Physical
Physical
TEL
NET
DNS SNMP NFS HTTP
RPC
TCP
Transport
Transport
Network
Network
L2
TCP/IP Protocol Suite
Internet
Internet
Protocol
Protocol
(IP)
(IP)
Protocol
Protocolfor
for
Underlying
Underlying
Network
Network
ICMP
IGMP
Ethernet
UDP
IP
Token
Ring
ARP
ATM
RARP
...
[Forouzan 2003]
Layer-by-Layer (OSI Model) View
sender
receiver
MS Outlook
Netscape
L7
through
L5
Hello!
Hello!
Encoded Data
L4
H4 Data1
H4 Data2
H4 Data3
L3
H3 Data1
H3 Other
H3 Data2
L2
H2 Data11 H2 Other
H2 Data12
L1
001100110000111100001111000
Encoded Data
packets
frames
H4 Data1
H4 Data2
H4 Data3
H3 Data1
H3 Other
H3 Data2
H2 Data11 H2 Other
H2 Data12
001100110000111100001111000
Activities Required (Sender Side)


Edit message and enter “send”  MS Outlook Express
Convert into sequence of bits

Tags must be inserted so that original message can be reconstructed at
destination




Encrypt message if necessary  for privacy
Compress if necessary
Partition into packets of fixed maximum size



E.g., “string” 01001000 … “JPEG” 110011101010 … “end”
 11001100100010 … 101011111100 … 01111110
Attach header information (Packet ID, destination, checksum, …)
IP address
Intersperse with packets from messages created by other applications
On first link of path,


Partition each packet into fixed-size frames (with headers)
Send each frame out onto the network
Activities Required on Network


Route each packet to its destination
During each “hop” of the path

Send signals back and forth to coordinate the sending and receiving of
the stream of bits corresponding to a frame


Check each frame for errors




Handshaking
Request retransmission in the case of errors
Arrange received frames into the proper order
Wait for all frames of the packet to be received
Once each packet reaches its destination node,


IP address
Store packet in a memory buffer at destination
Send signal to destination CPU to inform it of the arrival of the new
packet
Port Number
Activities at Destination Node

Receive packets








Check each packet for errors and request retransmission in the case of
errors
Arrange received packets into the proper order
Once all packets have been received, form a complete message
Decompress if necessary
Decrypt if necessary
Check for errors
Use tags in the bit stream to reconstruct the message
Show message to user using email tool (e.g., MS Outlook
Express)
Network Addresses

IP (Internet Protocol) address




MAC (Medium Access Control) address




Address used to identify a computing node on the internet
Network layer (L3) address
E.g., 141.223.165.189 (Look up “properties” on “TCP/IP” on “Network”)
Address used to identify a LAN card – cannot be changed
Data link layer (L2) address
E.g., abcd1234 (Enter “ipconfig /all” from MS Windows “cmd” window)
Port address


Address used to identify a network interface point for an application prog.
Corresponds to a memory buffer



Send a message - write to a memory buffer on a remote computer
Receive a message – read from a memory buffer on the local computer
Example: 39 (for FTP), 3000 (for a user-defined port)
Connection-Oriented and Connectionless
Networking

Connection-oriented networking

Uses a specific network path that is established for the duration of a
connection



Main advantage: reliable communication
Main implementation method: TCP (transfer control protocol)


Three phases: connection establishment, data transfer, connection termination
Used in the “parallel merge sort” socket-based program (TCP sockets interface)
Connectionless networking



Finds a new path for each packet sent
Main advantage: fast communication for short messages
Main implementation method: UDP (user datagram protocol)
Communication Performance Parameters (1)

Throughput (데이터 처리량)

Actual number of bits transmitted per second


Note 1: different from latency (지연시간)
Note 2: different from bandwidth (대역폭)
Most important communication performance parameter
 Typical measurement method


Send a data file from a source node to a destination node




Record the time t1 when the first byte of the data is received
Record the time t2 when the last byte of the data is received
Divide amount of data received by (t2 – t1)
Note: Mbps = mega-bits-per-second (not bytes)
Communication Performance Parameters (2)

Bandwidth

Maximum number of bits that can be transmitted per second


Note 1: different from latency (지연 시간)
Note 2: different from throughput (데이터 처리량)
Measures performance of network only (not the computer
hardware or software)
 Typical measurement method



Difficult to measure since effects of small data amounts, software and
hardware at source and destination nodes must be removed
The “rated” figure stated in the specifications for the relevant
communication protocol is most commonly used

E.g., 11 Mbps for IEEE 802.11b
Communication Performance Parameters (3)

Latency
Time required for the first byte of a message to be
transferred from the source to the destination node
 Should include software processing time
 Typical measurement method






At time t1, source node sends a very small message to destination node
Destination node receives message and sends it back to the source
node
Source node receives message and records the time t2
One-way communication latency is (t2 – t1) / 2
Why can’t we measure latency directly (record time t3 at destination
and measure latency as t3 – t1)?
Computer Communication Example
(Revisited)

Send picture image and message to friend
sender
Microsoft Outlook
“zero
copy”
[IBM’08]
Hello!
User
Memory
Space
OS Kernel
Memory
Space
receiver
1. Polling
2. Interrupt
3. DMA
Hello!
Netscape Messenger
Hello!
system software
system software
NIC H/W
(LAN card)
NIC
H/W
Section 7.8 of [Culler 1999]
17

Communication Microbenchmarks at 3 levels




Basic network transaction
Shared address space
Message passing using MPI
Why must this
type of echo
test be used?
Network Transaction Performance

Echo test using Active Messages (AM) user-level software
network interface
source
Send message;
receive reply;
compute 1-way
communication delay
k-byte message
destination
k-byte message
EECE 550
Receive message
and immediately
send reply
LogP Communication Model
18

LogP model used for network transaction performance
modeling





L  latency (within the physical network)
o  overhead (= sending overhead + receiving overhead)
g  gap (the minimum gap between consecutive message send
operations)
P  processing time (for normal processing of application
programs)
Refer to Figs. 7.30 and 7.31 [Culler 1999]
EECE 550
19

Message-Passing Operations

Simple model for overall time to send n bytes
 T(n) = T0 + n/B



T0 is time to send initial byte of data over the network
 Sending overhead + receiving overhead
n is number of bytes
B is the bandwidth of the network link
 rinfinity
: asymptotic bandwidth
 n½ : transfer size at which throughput = ½ * rinfinity
EECE 550

Table 7.1 of [Culler 1999]: progressive
improvement in T0, B, MFLOPS/processor
 Berkeley
NOW
 T0
= 6 microseconds
 rinfinity = 120 MB/s (Megabytes per second)
EECE 550
20
Application-Level Performance
21

How does LogP affect application performance?



Depends on the characteristics of the application
General trends observable
Figures 7.35, 7.36, 7.37, 7.38 and Table 7.2 [Culler 1999]
large  larger messages are preferable
 T0 small, B large  small messages are acceptable
 Larger numbers of processors  smaller message sizes,
smaller working sets (size of data that fits into faster
memory, such as one cache line)
 T0
EECE 550
Synchronization Issues
22

Message-Passing Model

Locks are not necessary since mutual exclusion is not a problem



Each process has exclusive access to its local memory and uses messagepassing to send/receive data from/to other nodes
Group synchronization and group communication is still a problem
Shared-Address-Space Model


Requires basic support for “locks” and “barriers”
Software algorithms execute on top of basic atomic exchange primitives

Programming environment/hardware must provide perception of atomic
memory operations
EECE 550
Group Communication Operations
23





Unicast (one-to-one)
Multicast (one-to-many)
Broadcast (one-to-all)
All-to-all broadcast
All-to-all personalized multicast (or broadcast)


Also referred to as “gossiping”
Special operations used for performance improvement


Parallel prefix (used with parallel supercomputers)
Map-reduce (white paper written by Google engineers)
EECE 550
Communication Support in the ESA Lab
Cluster


1Gbps Ethernet cards and switches
Myrinet switches, Myrinet LAN cards (from Myricom)



1.28 Gbps/port
TCP/IP, Myrinet GM and BIP LAN interface software [Kim 2001]
Myrinet2000 switch and Myrinet2000 LAN cards


2.0 Gbps/port bandwidth (= 250MBps)
TCP sockets


Myrinet GM LAN interface software (www.myricom.com)


> 100 microsecond latency, much less than peak BW
Around 5 microsecond latency, close to peak BW
Note: current (2009) state-of-art is Myrinet10G, MX S/W

Around 2 microsecond latency, close to 10Gbps throughput
References




Behrouz A. Forouzan, TCP/IP Protocol Suite, 2nd Ed.,
McGraw-Hill, Boston, 2003.
D. E. Culler, J. P. Singh and A. Gupta, Parallel Computer
Architecture: A Hardware/Software Approach, Morgan
Kaufmann, San Francisco, 1999.
http://www.ibm.com/developerworks/linux/library/jzerocopy/ , 2008.
S. C. Kim and S. Lee, ``Measurement and prediction of
communication latencies in Myrinet networks,'' J. Parallel
and Distributed Computing, Vol. 61, No. 11, pp. 16921704, November 2001.