Dr. Johnnie Baker, Dr. Robert Walker, and Dr. Jerry Potter (Emeritus),

November 18, 2005 PACL and ASC Processor Research Overview

1

Research Overview Parallel and Associative Computing Groupand theASC Processor Group

Kent State University

Dr. Johnnie Baker, Dr. Robert Walker, and Dr. Jerry Potter (Emeritus), Michael Scherger, Wittaya Chantamas, Hong Wang

Sabegh Singh Virdi, Shannon Steinfadt, Kevin Schaffer

Department of Computer Science


Kent, Ohio

PACL and ASC Processor Research Overview

2November 18, 2005

Presentation Outline

Associative and Parallel Algorithms and Applications Design of an Associative Parallel Database Server –

Michael Scherger Longest Common Subsequence Algorithm on ASC

Processors Using a Coterie Network – Sabegh Singh Virdi

Shannon’s Presentation Title Goes Here… Multithreaded ASC – Kevin Schaffer


3November 18, 2005

MASC Parallel Database Server

PE

…

PE

PE

PE

PE

PE

…

PE

PE

PE

PE

ASHDGALTLONLATAIDRBusy / Idle

ASHDGALTLONLATAIDRBusy / Idle

330907084.1339.54CO2341TT

……………………

F

F

F

F

30531529082.5340.0AA1223FT

525015081.5141.24UA722FT

2752256084.3139.9AA123FT

450027081.2640.55CO56TT

330907084.1339.54CO2341TT

……………………

F

F

F

F

30531529082.5340.0AA1223FT

525015081.5141.24UA722FT

2752256084.3139.9AA123FT

450027081.2640.55CO56TT


4November 18, 2005


Database is resident in parallel array memory.

Parallel memory map is similar to table layout.

Size of parallel array is fixed during execution. Only used physical PEs.

PE memory size is fixed.

Table records are placed in the array memory in “folds”.

Parallel array memory is allocated dynamically.

Coalescing parallel memory manager.

Tables will also have a set of parallel control bits.


5November 18, 2005


Table A1 Table A2

Database

Table List

Fold ListField List

Light shaded regions are empty records for table A

Dark shaded regions are

occupied records for table A

Stores relative

offsets of table fields

Stores base memory address

of each fold


6November 18, 2005

Multiple Tables and Folds

Table A1 Table B1 Table B2 Free Parallel Memory


7November 18, 2005

Parallel Database Operations

SQL INSERT Algorithm Perform associative search on

BI field to find first/next cell for insert.

If none found, create new fold.

Copy data from sequential to parallel memory.

O(1)…(O(#folds)) Uses broadcast / reduction

network.

Table A1 Table A2

INSERT INTO FLIGHTS( AID, LAT, LON, ALT, AS )

VALUES( ‘CO128’, 43.39, 83.67, 190, 450 )


8November 18, 2005

Parallel Database Operations

SQL DELETE Algorithm Perform associative search on

WHERE clause. (Flag responders)

Reset BI field where responder flag is set true. Data remains in memory.

If all PE records are idle, CPMM marks block as free.

O(1)…(O(#folds))

Table A1 Table A2

DELETE FROM FLIGHTS

WHERE AID = ‘NW 545’


9November 18, 2005

Future Work

MASC Parallel Database Server Change the parallel memory manager to support

multiple instruction streams

Table ATable BTable A x B

Free Parallel Memory

Internally Fragmented Parallel Memory


10

Longest Common Subsequence Algorithm on ASC Processors using

Coterie Network

Sabegh Singh Virdi

ASC Processor GroupComputer Science Department



11November 18, 2005


Introduction to String matching and its variations Role of LCS in Molecular Biology Overview of LCS Brief introduction on Coterie Network Longest Common Subsequence on Coterie Network

Exact match Approximate match

Summary and Future work


12November 18, 2005

String Matching

One of the most fundamental operation in computing.

Comparing two linear arrays of character Application in bioinformatics, searching

genetic databases String involved are how ever enormous,

efficient string processing is therefore a requirement


13November 18, 2005

String Matching Variations

Is Exact match the only solution? What if the pattern does not occur in the text? It still makes sense to find the longest

subsequence that occurs both in the pattern and in the text. This is the longest common subsequence problem

Longest Common Subsequence, Longest Common Substring, Sequence alignment, Edit distance Problem are all variation of SM problem


14November 18, 2005


String matching and its variations Role of LCS in Molecular Biology Overview of LCS Brief introduction on Coterie Network Longest Common Subsequence on Coterie Network




15November 18, 2005

Role of LCS in Molecular biology

DNA sequences (genes) can be represented as sequences of four letters A, C, G, and T corresponding to the four submolecules forming DNA

When biologists find a new sequences, they typically want to know what other sequences it is most similar to

One way of computing how similar (homologous) two sequences are, is to find the length of their longest common subsequence


16November 18, 2005

Role of LCS in Molecular biology

This is a simplification, since in the biological situation one would typically take into account not only the length of the LCS, but also e.g. how gaps occur when the LCS is embedded in the two original sequences.

An obvious measure for the closeness of two strings is to find the maximum number of identical symbols (preserving symbol order)

This by definition, is the longest common subsequence of the strings


17November 18, 2005

Overview of LCS Algorithm

Given two strings, find the LCS common to both strings. Example:

String 1: AGACTGAGGTA String 2: ACTGAG

AGACTGAGGTA - -ACTGAG - - - list of possible alignments - -ACTGA - G- - A- -CTGA - G- - A- -CTGAG - - -

The time complexity of this algorithm is clearly O(nm);


18November 18, 2005


Actually this time does not depend on the sequences u and v themselves but only on their lengths

The bottleneck in efficient parallelization of LCS problem are the calculating the value of diagonal elements, as shown

As seen, the value of {i,j} depend upon the previous element {i-1,j-1}, when a match is found


19November 18, 2005

Possibility of more then one LCS Associate some parameters Take account into gap penalties The Smith-Waterman Algorithm uses the

same concept that of LCS algorithm, but gives us the optimal result



20November 18, 2005


1 1 1 1 1

11

2111

1 222222

111111

3

1

1

1

44443222

3333

43332

5

55

43332 6

5

4

3

2 2

666

5 5

4

3

0 0 0 0 0 0 0 0 0 0 0 0

A G A C T G A G G T A

0

0

0

0

0

0

A

C

T

G

A

G


21November 18, 2005

Communication between PE’s

In 2D mesh network, Communication between P.E’s themselves take place

in two different ways By using the nearest neighbors mesh

interconnection network Powerful variation on the nearest-neighbor mesh

called the “Coterie network”, developed in response to the requirement for nonlocal communication

Properties significantly different from the usual mesh


22November 18, 2005






23November 18, 2005

Coteries[ Weems & Herbordt ]

“A small often selected group of persons who associate with one another frequently” Features:

Related to other Reconfigurable broadcast network Describable using hypergraphs Dynamic in nature

Advantages: Propagation of information quickly over long distances at

electrical speed Support of one-to-many communication within coterie,

reconfigurability of the coterie


24November 18, 2005

PE’s form Coteries

5 x 5 coterie network with switches shown in “arbitrary” settings. Shaded areas denotes coterie (the set of PEs Sharing same circuit)


25November 18, 2005

Coterie’s Physical Structure

In the Physical implementation, each PE controls set of switches Four of these switches control

access in the different directions (N,S,E,W)

Two switches H and V are used to emulated horizontal and vertical buses

The last two switches NE and NW are used to creation of eight way connected region

NWNE

WSES

V

H E

S

W

: Switch

N


26November 18, 2005






27November 18, 2005

LCS Algorithm on Coterie Network



28November 18, 2005




29November 18, 2005








Content of each PE’s after MULTICAST operation


30November 18, 2005


A

C

T

G

A

G


31November 18, 2005


A

C

T

G

A

G


32November 18, 2005


A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

A

C

T

G

A

G

Content of each PE’s after MULTICAST operation


33November 18, 2005


1 0 1 0 0

00

0000

0 000001

100010

1

0

0

1

11010001

0000

00010

0

01

10001 1

0

0

1

0 0

001

0 1

0

0


A

C

T

G

A

G


34November 18, 2005


1 0 1 0 0

00

0000

0 000001

100010

1

0

0

1

11010001

0000

00010

0

01

10001 1

0

0

1

0 0

001

0 1

0

0


A

C

T

G

A

G

Inject unique token


35November 18, 2005


We try to refine the Exact Match algorithm to support approximate matching

We make use of tokens The next example demonstrate this

problemFor the string:

Text :AGACTGAGGTA Pattern : ACTAAG


36November 18, 2005






37November 18, 2005


1 0 1 0 0

00

0000

0 000001

100010

1

1

0

1

00100010

0000

00010

0

01

10001 1

0

0

1

0 0

001

0 1

1

0


A

C

T

A

A

G

Inject unique token


38November 18, 2005

Token method

In this method, we explicitly close the W-S switch based on some condition

Inject unique token symbols Where the H and V switch is set within a PE’s,

we close the W-S switch as shown As the token traverse, keep track of gaps and

match Resulting a path from first row to the last row Value in CR gives the length of LCS and in SR

number of gaps occurred, first row of PE


39November 18, 2005


1 0 1 0 0

00

0000

0 000001

100010

1

1

0

1

00100010

0000

00010

0

01

10001 1

0

0

1

0 0

001

0 1

1

0


A

C

T

A

A

G

Inject unique token


40November 18, 2005






41November 18, 2005


We have presented two variation of the lcs algorithm

We have Explored a new network for this problem

Constant time algorithm for Exact matchApproximate algorithm depends upon

the diameter of the network


42November 18, 2005


Future Work:

Optimize the algorithm for Approximate match Implementing the algorithm on FPGA’s model Incorporating the Don’t Care Symbol Extend the idea to support sequence

alignmentConserve memory by using encoding schemeWe can use Virtual simulation of PEs, in case

we ran out of PEs


43November 18, 2005

Acknowledgements

Professor Walker Professor Baker Professor Weems Professor Herbordt Professor Piontkivska Kevin Schaffer, Hong Wang, Shannon Steinfadt,

Jalpesh Chitalia, and Michael Scherger


44November 18, 2005

THANK YOU


45

Multithreaded ASC

Kevin Schaffer

ASC Processor GroupComputer Science Department



46November 18, 2005

Motivation

As we scale up the number of PEs, broadcast and reduction operations become more expensive due to wire delays

Deeper pipelines can absorb wire delays, but the cost is higher operational latencies

It becomes more difficult to keep the pipeline fully utilized, especially in associative programs Maximum/minimum search Branch if any responders

Possible solution: use instructions from multiple threads to keep the pipeline full


47November 18, 2005

Execution Time vs. Latency

0

50

100

150

200

250

300

350

400

450

1 2 3 4 5 6 7 8

Communication Latency (cycles)

Execu

tio

n T

ime (

cycle

s)

ASC Multithreaded ASC MASC


48November 18, 2005

Throughput vs. Latency

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

1 2 3 4 5 6 7 8

Communication Latency (cycles)

No

rmalized

Th

rou

gh

pu

t (i

nstr

ucti

on

s/c

ycle

)

ASC Multithreaded ASC MASC


49November 18, 2005

Designing a SIMD Pipeline

Separate scalar and parallel pipelines Parallel pipeline includes extra stages to handle

broadcast and reduction delays Allows scalar and parallel instructions to complete out of

order with respect to each other

I F I D B 1 R F

E X

W B

M

E X M R 4R 3B 2

W B


50November 18, 2005

Broadcast Dependency

Due to the shorter scalar pipeline, forwarding can eliminate stalls caused by broadcast dependencies

A longer scalar pipeline would have caused a stall

IF ID E X W B

1 2 3 4 5

ADD S2, S3, S4

PSUB P5, P6, S2

6

M

IF ID B1 RFB2 EX M R3 WBR4

PAND P7, P8, S2 IF ID B 1 R FB 2 E X M R 3 W BR 4

7 8 9 10 11 12


51November 18, 2005

Reduction Dependency

No amount of forwarding hardware can eliminate stalls due to reduction dependencies

Must use instructions from other threads to fill in the unused execution slots

IF ID B 1 R F

IF ID E X W B

1 2 3 4 5

RMAX S2, P3

ADD S5, S6, S2 ID

6

B2

ID

7

R1 R2 R3 WBR4

ID ID ID ID

8 9 10 11


52November 18, 2005

Instruction Issue

Use instructions from multiple threads to keep the pipeline fully utilized

Instructions within a thread are issued in order Scalar and parallel instruction may complete out of

order, so we check for hazards before issue Fine-grain multithreading interleaves threads at the

instruction level For now, at most one instruction issues each cycle With separate scalar and parallel pipelines, we could

issue one instruction to each


53November 18, 2005

Working with Threads

Management Fork and join instructions start and end threads Threads are allocated dynamically by the hardware

Communication Send and receive instructions can transmit small

amounts of data quickly Used to synchronize register contents between parent

and child threads Can use shared memory for large data


54November 18, 2005

Multithreaded ASC and MASC

All processors execute the same instruction at the same time in lock step

Each thread can use all the processors Multiple threads improve pipeline utilization, but

not processor utilization Multithreading and MASC can be used together

in order to improve both

Documents

Dr. Johnnie Baker, Dr. Robert Walker, and Dr. Jerry Potter (Emeritus),