Intel NAS Performance Toolkit - SNIA€¦ · Agenda Home/SMB Performance NAS Performance Toolkit Introduction Consumer NAS Performance Overview Using NASPT to improve NAS Performance

Storage Developer Conference 2008 Copyright © 2008, Intel Corporation. All rights reserved.

www.storage-developer.org

Intel® NAS Performance Toolkit

Anthony Bock

Frank Hady

Storage Technology Group, Intel Corporation



Agenda

Home/SMB PerformanceNAS Performance Toolkit IntroductionConsumer NAS Performance OverviewUsing NASPT to improve NAS PerformanceNASPT v1.7 Announcement

*Other names and brands may be claimed as the property of others. Intel is a trademark of Intel Corporation in the U.S. and other countries.



Role of DH/SMB NAS Evolving

More Files/More DataExpanding Usage ModelsNew Client Platforms

Laptops with SSDsMIDsNettops/NetbooksConnected TVs

Expanding Role for Consumer NAS Devices

AFP

CIFS/SMB

NFS

Media DistributionSearch

Backup

HTTP

UPnP

FTP

File Sharing

MetadataManagement



NASPT Designed Specifically for Consumer NAS Requirements

Existing Tools Repurposed from other Uses

Local Storage/Network Tools

Commercial Desktop Benchmarks

Enterprise Storage Tools

Need: Model Wide Range of Consumer NAS Workloads

Useful for Developers

Compelling to Consumers



Performance Tool Requirements

Easy to UseMinimal Learning CurveEasily configured and runNo special test hardware

AccurateMinimal test client impactReproducible measurementsComparable across NAS devices

End UserRelevant

Usage derived Test CasesEasy to Understand Measurements



Agenda




Intel NAS Performance Toolkit

Models real applications

Trace based workloads

Enables file level buffering

Writes to both existing and new files

Controls for:

Drive layout

Background processes

Includes graphical analyzer



NASPT Based on Real WorkloadsSe

rvin

g T

wo

Vid

eo S

trea

ms



NASPT Based on Real Workloads

Cop

ying

a W

hole

Dir

ecto

ry



NASPT Based on Real WorkloadsBr

owsi

ng a

Col

lect

ion

of P

hoto

s



Application Based Workloads

Test # files % seq. Description

HD Video Play 1 99.5% 720p HD stream from Windows Media Player* 256kB reads

2HD Video Play 2 18.1% 2x playback

4HD Video Play 4 9.6% 4x playback

HD Video Record 1 99.9% 720p HD stream, 256kB writes

HD Video Play & Record 2 17.8% 1 playback, 1 record simultaneously

Directory Copy From NAS 2833 52.5% 64kB reads

Directory Copy To NAS 2833 52.5% Predominantly 64kB writes, wide scattering under 16kB

File Copy From NAS 1 100% 4GB file copy, 64kB reads

File Copy To NAS 1 100% 64kB writes

Photo Album 169 80% All reads – wide distribution of sizes

Office Productivity 607 81.3% Reads & writes; small, 1kB & 4kB reads; Mostly 1kB writes

Content Creation 98 38.6% 95% writes; 1k, 4k & little reads; Writes up to 64kB

More realistic workloads than synthetic tests



Write to New vs. Write to Existing

Overwriting an existing file and writing to a new file differ in performanceApplications do bothMany tools use just one method

To better model real apps, NASPT can use either approach as specified by the workload

Sample NASOverwrite vs. New Write

Source: NASPT Measurements



Disk Layout Matters to NAS Performance

Disk layout has significant impact on performance

Applications have no control over layout

Some tools use direct/ unbuffered I/O to force idealized layout

NASPT’s Batch Mode performs 5 trials with independent disk images

Result is median of all five trials

Retains output traces from all trials

0

5

10

15

20

25

30

35

HDVideo_

1Play

HDVideo_

2Play

HDVideo_

4Play

HDVideo_1

Record

HDVideo_

1Play

_1Rec

ordCon

tentC

reatio

nOffic

eProd

uctiv

ity

Thro

ughp

ut (M

B/s

)

Trial 1

Trial 2

Trial 3

Trial 4

Trial 5




Agenda




NAS Usage: Still & Streaming Media

Digital Photo Browsing(~10 minute fetch of 100+ images)

0 1 2 3 4

Storage Wait Time (Minutes)

Local HDFastest NASSlowest NAS

Interactive Usage Highlights Value of Performance

5.6x slower than local HD2.7x slower than fastest NAS

Video Distribution

0

10

20

30

40

50

60

1 Playback 2 Playback 1 PB + 1Record

4 Playback

Number of Independent Streams

Max

mim

um T

hrou

ghpu

t (M

B/s

)

Slowest NASFastest NAS

Source: NASPT MeasurementsSource: NASPT Measurements

1080i

1080p



Consumer NAS Performance Varies Widely

Be a NAS performance leader – Users will notice

6x 6x

12x

4x

5x

10x

Source: NASPT measurements

9x



Users Will Notice…

100Mb/s Ethernet

802.11g WiFi

802.11n WiFi

Gb Ethernet

Many users will already notice superior NAS performanceAs users upgrade networks, NAS performance becomes more and more visible


ITU G.hn target



Agenda




Realistic File Usage Enables NASPT to find Issues

NTFS and EXT3 Handle “Data Holes” Differently

ApplicationTwo writes to file “foo”

with never-written“data hole” between

NTFSGap is preserved as allocated

but invalid disk space

EXT3Gap is eliminated by abutting

discontiguous writes

21 21210 EOF

unwritten

And with “file buffering” on, Windows* generates 1byte hintsCombination Leads to Unintended Consequences for Linux* NAS

ApplicationContiguous write stream

EXT3 resultSevere fragmentation - ~18/MBLow performance – 50% drop

0 EOF

NetworkWindows OS adds one byte

writes to high offsets

Application Write - Contiguous

1-byte write – HIGH offset




1-byte write – HIGH offset

Tim

e

1-byte writes placed adjacent to app. writes

App writes and hints continue to

interleave

When app writes reach hint

offsets, ext3 tries to fill in



Samba’s included “strict allocate” feature fills gaps with zeroesDisk layout improves dramaticallyFiles with many discontinuous writes and large gaps experience severe delays

Strict Allocation Doesn’t Quite Work

ApplicationTwo writes to file “foo”

with never-written“data hole” between

Strict AllocateGap is preserved by filling

intervening space with zeroes

000000000 21210 EOF

unwritten



Solution: Modify Samba Zero Fills

Two small changes improve strict allocate behavior with Windows* clientsSkip zero fills when copying files

1-byte hints don’t occur when file size is known ahead of timeOnly fill to 2MB past current end of file

0

10

20

30

40

50

60

Thro

ughp

ut (M

B/s

)

HDVideo_1

PlayHDVideo

_2Play

HDVideo_3

PlayHDVideo

_4Play

HDVideo_1

Record

HDVideo_1

Play_1

Record

ContentC

reation

DefaultStrict AllocateSA Fix

samba/source/smbd/vfs.cvfs_fill_sparse(…){

…

if (len <= st.st_size) return 0;

//Impose limit on how much to write ahead of current position#define ALLOCATION_LIMIT 0x200000if (len - st.st_size > ALLOCATION_LIMIT)

return 0;

…}

samba/source/modules/vfs_default.c:vfswrap_ftruncate(…){

int result = -1;SMB_STRUCT_STAT st;char c = 0;SMB_OFF_T currpos;

START_PROFILE(syscall_ftruncate);

/* ignore file fill when presented with new file of known size.if (lp_strict_allocate(SNUM(fsp->conn))){

result = strict_allocate_ftruncate(handle, fsp, fd, len);END_PROFILE(syscall_ftruncate);return result;

}*/…

}



Agenda




NASPT Version 1.7 Available

General usability enhancementsReduced run timeBetter controls for sources of variance

Variations in disk layoutCompetition from background processes

Users may add custom workloads

http://www.intel.com/software/naspthttp://www.intel.com/design/servers/storage/NAS_Perf_Toolkit.htm



Q & A

Would this tool be valuable to you as an open source project?How many would anticipate contributing code?

Intel NAS Performance Toolkit - SNIA€¦ · Agenda Home/SMB Performance NAS Performance Toolkit Introduction Consumer NAS Performance Overview Using NASPT to improve NAS Performance

Documents