EMC PresentationApril 20051 Research @ Northeastern University I/O storage modeling and performance –David Kaeli Soft error modeling and mitigation –Mehdi

EMC Presentation April 2005 1

Research @ Northeastern University

• I/O storage modeling and performance – David Kaeli

• Soft error modeling and mitigation – Mehdi B. Tahoori

I/O Storage Research at Northeastern University

David Kaeli

Yijian WangDepartment of Electrical and Computer Engineering

Northeastern University

Boston, MA

[email protected]


Outline

• Motivation to study file-based I/O

• Profile-driven partitioning for parallel file I/O

• I/O Qualification Laboratory @ NU

• Areas for future work


Important File-base I/O Workloads

• Many subsurface sensing and imaging workloads involve file-based I/O– Cellular biology – in-vitro fertilization with NU biologists– Medical imaging – cancer therapy with MGH– Underwater mapping – multi-sensor fusion with Woods Hole

Oceanographic Institution– Ground-penetrating radar – toxic waste tracking with Idaho

National Labs


The Impact of Profile-guided Parallelization on SSI Applications

• Reduced the runtime of a single-body Steepest Descent Fast Multipole Method (SDFMM) application by 74% on a 32-node Beowulf cluster

• Hot-path parallelization• Data restructuring

• Reduced the runtime of a Monte Carloscattered light simulation by 98% on a 16-node Silicon Graphics Origin 2000

• Matlab-to-C compliation• Hot-path parallelization

• Obtained superlinear speedup of Ellipsoid Algorithm run on a 16-node IBM SP2

• Matlab-to-C compliation• Hot-path parallelization

Soil

Air

Mine

Scattered Light Simulation Speedup

1

10

100

1000

10000

100000

Ru

nti

me

in s

ec

on

ds

Original

Matlab-to-C

Hot pathparallelization

Ellipsoid Algorithm Speedup(versus serial C version)

05

101520

1 2 4 8 16

Number of Nodes

Sp

ee

du

p

64-vector 256-vector1024-vector linear speedup


Limits of Parallelization

• For compute-bound workloads, Beowulf clusters can be used effectively to overcome computational barriers

• Middlewares (e.g., MPI and MPI/IO) can significantly reduce the programming effort on parallel systems

• Multiple clusters can be combined, utilizing Grid Middleware (Globus Toolkit)

• For file-based I/O-bound workloads, Beowulf clusters and Grid systems are presently ill-suited to exploit the potential parallelism present on these systems


Outline






Parallel I/O Acceleration• The I/O bottleneck

– The growing gap between the speed of processors, networks and underlying I/O devices

– Many imaging and scientific applications access disks very frequently

• I/O intensive applications– Out-of-core applications

– Work on large datasets that cannot fit in main memory

– File-intensive applications– Access file-based datasets frequently– Large number of file operations


Introduction

• Storage architectures– Direct Attached Storage (DAS)

– Storage device is directly attached to the computer

– Network Attached Storage (NAS)– Storage subsystem is attached to a network of servers

and file requests are passed through a parallel filesystem to the centralized storage device

– Storage Area Network (SAN)– A dedicated network to provide an any-to-any connection

between processors and disks


I/O PartitioningP

An I/O intensive application

Disk

P P P…

Disk

Disk Disk Disk

P P P…

…

Data Partitioning

Multiple disks(i.e. RAID)

Disk Disk Disk

P

…Data Striping

Multiple Processes

(i.e. MPI-IO)


I/O Partitioning• I/O is parallelized at both the application level

(using MPI and MPI-IO) and the disk level (using file partitioning)

• Ideally, every process will only access files on local disk (though this is typically not possible due to data sharing)

• How to recognize the access patterns?• Profile-guided approach


Profile Generation

Run the application

Capture I/O execution profiles

Apply our partitioning algorithm

Rerun the tuned application


I/O traces and partitioning• For every process, for every contiguous file access,

we capture the following I/O profile information:– Process ID– File ID– Address– Chunk size– I/O operation (read/write)– Timestamp

• Generate a partition for every process• Optimal partitioning is NP-complete, so we develop a

greedy algorithm• We have found we can use partial profiles to guide

partitioning


for each IO process, create a partition;

for each contiguous data chunk {

total up the # of read/write accesses on a process-ID basis;

if the chunk is accessed by only one process

assign the chunk to the associated partition;

if the chunk is read (but never written) by multiple processes

duplicate the chunk in all partitions where read;

if the chunk is written by one process, but later read by multiple {

assign the chunk to all partitions where read and broadcast the updates on writes;

else assign the chunk to a shared partition;

} }

For each partition

sort chunks based on the earliest timestamp for each chunk;

Greedy File Partitioning Algorithm


Parallel I/O Workloads• NASA Parallel Benchmark (NPB2.4)/BT

– Computational fluid dynamics– Generates a file (~1.6 GB) dynamically and then reads it back– Writes/reads sequentially in chunk sizes of 2040 Bytes

• SPEChpc96/seismic– Seismic processing– Generates a file (~1.5 GB) dynamically and then reads it back– Writes sequential chunks of 96 KB and reads sequential chunks of 2 KB

• Tile-IO– Parallel Benchmarking Consortium– Tile access to a two-dimensional matrix (~1 GB) with overlap– Writes/reads sequential chunks of 32 KB, with 2KB of overlap

• Perf– Parallel I/O test program within MPICH– Writes a 1 MB chunk at a location determined by rank, no overlap

• Mandelbrot– An image processing application that includes visualization– Chunk size is dependent on the number of processes


10/100Mb Ethernet Switch

RAIDNode

LocalPCI-IDE

Disk

LocalPCI-IDE

Disk

P2-350Mhz

P2-350Mhz P2-350Mhz

P2-350Mhz

P2-350Mhz

RAIDNode

P2-350Mhz

Beowulf Cluster


Hardware Specifics• DAS configuration

– Linux box, Western Digital WD800BB (IDE), 80GB, 7200RPM

• Beowulf cluster (base configuration)– Fast Ethernet 100Mbits/sec– Network Attached RAID - Morstor TF200 with 6-9GB drives

Seagate SCSI disks, 7200rpm, RAID-5 – Local attached IDE disks – IBM UltraATA-350840, 5400rpm

• Fibre channel disks– Seagate Cheetah X15 ST-336752FC, 15000rpm


0

50

100

150

200

UnixWrite

UnixRead

MPI-IOWrite

MPI-IORead

P-IOWrite

P-IORead

Ban

dwid

th (

MB/s

ec)

4 procs9 procs16 procs25 procs

Write/Read Bandwidth

0

50

100

150

200

UnixWrite

UnixRead

MPI-IOWrite

MPI-IORead

P-IOWrite

P-IORead

Ban

dwid

th (

MB/s

ec)

4 procs8 procs16 procs24 procs

NPB2.4/BT

SPECHPC/seis


0

25

50

75

100

125

MPI write MPI read PIO write PIO read

Band

wid

th (

MB/

sec)

Write/Read Bandwidth

0

50

100

150

200

250


Band

wid

th (

MB/

sec)

0

50

100

150

200

250


Bandw

idth

(M

B/se

c)

4 procs8 procs

16 procs24 procs

MPI-Tile Perf

Mandelbrot


Total Execution Time

0

1000

2000

3000

4000

Ex

ec

uti

on

Tim

e (

se

co

nd

s)

MPI-IO

PIO


Profile training sensitivity analysis

• We have found that IO access patterns are independent of file-based data values

• When we increase the problem size or reduce the number of processes, either:– the number of IOs increases, but access patterns

and chunk size remain the same (SPEChpc96, Mandelbrot), or

– the number of IOs and IO access patterns remain the same, but the chunk size increases (NBT, Tile-IO, Perf)

• Re-profiling can be avoided


Execution-driven Parallel I/O Modeling

• Growing need to process large, complex datasets in high performance parallel computing applications

• Efficient implementation of storage architectures can significantly improve system performance

• An accurate simulation environment for users to test and evaluate different storage architectures and applications


Execution-driven I/O Modeling • Target applications: parallel scientific programs

(MPI)• Target machine/Host machine: Beowulf clusters• Use DiskSim as the underlying disk drive

simulator• Direct execution to model CPU and network

communication• We execute the real parallel I/O accesses and

meanwhile, calculate the simulated I/O response time


Validation – Synthetic I/O Workload on DAS

Response Time of Sequential Writes

0

2

4

6

8

10

12

1 2 4 8 16access size in number of blocks

number of accesses = 1000

seco

nd

s

Response Time of Sequential Reads

0

2

4

6

8

10

1 2 4 8 16access size in number of blocks

number of accesses = 1000

seco

nd

s

modelreal

Response Time of Non-contiguous Reads

0

2

4

6

8

10

1 2 4 8 16 32seek distance in number of blocks

access size = 1 blocknumber of accesses = 1000

seco

nds

Response Time of Non-contiguous Writes

0

2

4

6

8

10

1 2 4 8 16 32seek distance in number of blocks

access size = 1 blocknumber of accesses = 1000

seco

nd

s


Simulation Framework - NAS

LAN/WAN

Network File System

I/O traces

Local I/O traces Local I/O traces Local I/O traces Local I/O traces

RAID controller

DiskSim

I/O requests

Filesystem metadata

Logical file access addresses


Execution Time of NPB2.4/BT on NAS - base configuration

0

500

1000

1500

2000

2500

3000

3500

4000

4 9 16 25number of processors

seco

nds

model

real


LAN/WAN

FileSystem FileSystem FileSystem FileSystem

I/O traces I/O traces I/O traces I/O traces

DiskSim

DiskSim

DiskSim

DiskSim

Simulation Framework – SAN direct• A variety of SAN where disks are distributed across the network and each server is directly connected to a single device• File partitioning• Utilize I/O profiling and data partitioning heuristics to distribute portions of files to disks close to the processing nodes


Execution Time of NPB2.4/BT on SAN-direct - base configuration

0

500

1000

1500

2000

2500

3000

4 9 16 25

number of processors

seco

nds

model

real


Hardware Specifications


I/O Bandwidth of SPEChpc/seis

0

50

100

150

200

250

NA

S-jo

ulian

NA

S-A

TA

NA

S-S

CS

I

NA

S-F

C

SA

N-jo

ulian

SA

N-d

irect -AT

A

SA

N-d

irect-SC

SI

SA

N-d

irect-FC

storage architectures

MB

/s

4 processors

8 processors

16 processors


I/O Bandwidth of Mandelbrot

050

100150200250300350400

NA

S-jo

ulian

NA

S-A

TA

NA

S-S

CS

I

NA

S-F

C

SA

N-jo

ulian

SA

N-d

irect -AT

A

SA

N-d

irect-SC

SI

SA

N-d

irect-FC

storage architectures

MB

/s

4 processors

8 processors

16 processors


Publications• “Profile-guided File Partitioning on Beowulf Clusters,” Journal of Cluster

Computing, Special Issue on Parallel I/O, to appear 2005.• “Execution-Driven Simulation of Network Storage Systems,” Proceedings of the 12th

ACM/IEEE International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS), October 2004, pp. 604-611.

• “Profile-Guided I/O Partitioning,” Proceedings of the 17th ACM International Symposium on Supercomputing, June 2003, pp. 252-260.

• “Source Level Transformations to Apply I/O Data Partitioning,” Proceedings of the IEEE Workshop on Storage Network Architecture And Parallel IO, Oct. 2003, pp. 12-21.

• “Profile-Based Characterization and Tuning for Subsurface Sensing and Imaging Applications,” International Journal of Systems, Science and Technology, September 2002, pp. 40-55.


Summary of Cluster-based Work • Many imaging applications are dominated by file-based

I/O• Parallel systems can only be effectively utilized if I/O is

also parallelized • Developed a profile-guided approach to I/O data

partitioning• Impacting clinical trials at MGH• Reduced overall execution time by 27-82% over MPI-IO• Execution-driven I/O model is highly accurate and

provides significant modeling flexibility


Outline






I/O Qualification Laboratory

• Working with Enterprise Strategy Group• Develop a state-of-the-art facility to provide

independent performance qualification of Enterprise Storage systems

• Provide a quarterly report to ES customer base on the status of current ES offerings

• Work with leading ES vendors to provide them with custom early performance evaluation of their beta products



• Contacted by IOIntegrity and SANGATE for product qualification

• Developed potential partners that are leaders in the ES field

• Initial proposals already reviewed by IBM, Hitachi and other ES vendors

• Looking for initial endorsement from industry



• Why @ NU– Track record with industry (EMC, IBM,

Sun)– Experience with benchmarking and IO

characterization– Interesting set of applications (medical,

environmental, etc.)– Great opportunity to work within the

cooperative education model


Outline






Areas for Future Work

• Designing a Peer-to-Peer storage system on a Grid system by partitioning datasets across geographically distributed storage devices

joulian.hpcl.neu.edu keys.ece.neu.edu

InternetInternet

1Gbit/s100Mbit/sRAID

31 sub-nodes 8 sub-nodes

Head node Head node


NPB2.4/BT read performance

0

20

40

60

80

100

120

140

160

180

single server dual server P2P

MB/s

4 procs

9 procs

16 procs

25 procs


Areas for Future Work

• Reduce simulation time by identifying characteristic “phases” in I/O workloads

• Apply machine learning algorithms to identify clusters of representative I/O behavior

• Utilize K-Means and Multinomial clustering to obtain high fidelity in simulation runs utilizing sampled I/O behavior

“A Multinomial Clustering Model for Fast Simulation of Architecture Designs”, submitted to the 2005 ACM KDD Conference.

Documents

EMC PresentationApril 20051 Research @ Northeastern University I/O storage modeling and performance –David Kaeli Soft error modeling and mitigation –Mehdi