Tom Coughlan1

Trends in the Enterprise Storage Market

Tom CoughlanSr. Engineering Manager, Red HatFeburary, 2012

Tom Coughlan2

➔Big Data, ➔Unstructured Data, ➔Scale-Out vs. Scale-Up,➔Virtualization,➔pNFS➔Solid State Storage...

What is the future of SAN, NAS, DAS?

What role will Linux play in the new environment?

Hot Topics

Tom Coughlan3

The amount of data is exploding

● IBM estimates:● Every day, 2.5 exabytes of data are created.● 90% of the data in the world today was created within the past

two years.

● IDC projections:● transactional data will grow

at a 21.8% CAGR● unstructured data will grow

at a 61.7% CAGR

0 1 2 3 4 5 60

2

4

6

8

10

12

14

16

18

20

http://www-01.ibm.com/software/data/bigdata/http://searchstorage.techtarget.com/magazineContent/Object-storage-gains-steam-as-unstructured-data-grows

Tom Coughlan4

Much of this data is unstructured

● Total Archived Capacity, by Content Type, Worldwide, 2008-2015 (Petabytes) (ESG)

2008 2009 2010 2011 2012 2013 2014 20150

50000

100000

150000

200000

250000

300000

350000

FileE-mailDatabase

h20195.www2.hp.com/v2/GetPDF.aspx/4AA0-4382ENW.pdf

Tom Coughlan5

Big Data Analytics

● Healthcare, government, entertainment, social networking, oil and gas, retail...

● Business intelligence, strategy, product support, product panning, development, just-in-time capacity planning...

● Requires:● High volume (so cost control is critical)● Low latency (streaming)● Data integration: text, audio, video..., from retail,

medical, sensor, seismic, climate, satellite, (even databases)...

Tom Coughlan6

How to deal with this?

● So far, big data adherents have

1)Preferred to avoid shared storage, to minimize latency, and cost.

2)When more capacity is required, scale-out (add nodes)● Keep the storage close to the processor● Add processing power and storage together ● Use commodity parts● File replication for data persistence, as needed

ApplicationServer

ApplicationServer

ApplicationServer

IP Network

Tom Coughlan7

Scale-out, shared nothing

● Advantages:● Performs well for highly distributable problems● Inexpensive commodity hardware● Takes advantage of high performance local storage

(PCIe flash)

Tom Coughlan8

Scale-out, shared nothing

● Disadvantages:● Latency increases for queries that span nodes, and for

replication.● Specialized functions previously done in the storage

controller must be implemented in the o.s.: ● distributed fs, global namespace, data replication, backup,

encryption, snapshot, thin provisioning, remote replication, deduplication, compression, proactive error detection, ease of management, ...

Tom Coughlan9

Contrast with the more traditional approach- shared storage, scale-up

SAN or NAS

ApplicationServer

ApplicationServer

● Smaller number of nodes, more tightly-coupled, shared resources, specialized storage servers.

● When more capacity is required, scale-up existing nodes.

Ctrlr 1 Ctrlr 2 Ctrlr 1 Ctrlr 2

Tom Coughlan10

Scale up – add horsepower to existing nodes

● Advantages:● Some applications (single-threaded with large data sets)

can not be easily partitioned.● Centralized data protection, management, backup...

● Disadvantages:● Scaling limits... eventually you hit a wall

● ...and, if you do add another server+storage cluster, the lack of a global namespace can make it difficult to manage/load-balance the environment

● Proprietary, vendor lock-in● Generally more expensive

Tom Coughlan11

Shared storage

● Advantages:● Data is available to multiple machines

● Server virtualization provides load balancing

● Centralized data protection, management, backup...

● Disadvantages:● Access coordination can impact performance● Can be more expensive than DAS

Tom Coughlan12

As big data moves to the enterprise

● Take advantage of the scale-out approach● Control cost

● but keep shared storage● Virtualization● Ease-of-management, data protection, specialized

functions.

Tom Coughlan13

Scale-out, with shared storage

ApplicationServer

ApplicationServer

ApplicationServer

IP Network

NAS/SAN Shared Storage

● Scale-out NAS

● pNFS

● iSCSI and FCoE

Tom Coughlan14

Scale-out NAS (the hardware approach)

● Hardware vendors solve the storage controller bottleneck by “clustering” the controllers together.

● The group appears as one to the o.s.

ApplicationServer

ApplicationServer

ApplicationServer

IP Network

SAN or NAS

Ctrlr 1 Ctrlr 2 Ctrlr 1 Ctrlr 2 Ctrlr 1 Ctrlr 2

Tom Coughlan15

Scale-out NAS (the software approach) - Gluster Distributed Filesystem

IP NetworkGluster Global Namespace

Mirror Mirror

Distribute

Applicationw. Native GlusterFS

GlusterServer

xfs

GlusterServer

xfs

GlusterServer

xfs

GlusterServer

xfs

Applicationw. Native GlusterFS

Tom Coughlan16

Gluster with Non-native Clients

IP NetworkGluster Global Namespace

Application- NFS- CIFS

Mirror Mirror

Distribute

GlusterServer

xfs

GlusterServer

xfs

GlusterServer

xfs

GlusterServer

xfs

Applicationw. Native GlusterFS

Applicationw. Native GlusterFS

Tom Coughlan17

pNFS

● Access an NFS Metadata Server, then R/W the storage directly

● Data may be File, or Object, or Block Based

ApplicationServer

ApplicationServer

ApplicationServer

IP Network

SAN or NAS NFS metadata

Data

Tom Coughlan18

iSCSI and FCoE

● Lower-cost shared block storage● Traditional db, and virtualization workloads ● May pair nicely with pNFS:

ApplicationServer

ApplicationServer

ApplicationServer

IP Network

LANNFS metadata

Data

Tom Coughlan19

Conclusions

● In the strict big data approach, with no shared storage, the Linux system must perform the specialized functions previously performed by the storage controller.

● Currently underway:● efficient snapshot● thin provisioning● disk encryption - dm-crypt● integration of LVM with md RAID● PCIe flash performance optimizations

● Future:● hierarchical storage

Tom Coughlan20

Disk Density

30% CAGR

http://www.storagenewsletter.com/news/disk/hdd-technology-trends-ibm

Tom Coughlan21

Disk Capacity

http://en.wikipedia.org/wiki/File:Hard_drive_capacity_over_time.png

Tom Coughlan22

Disk Max. Sustained Bandwidth

10-15% CAGR

http://www.storagenewsletter.com/news/disk/hdd-technology-trends-ibm

Tom Coughlan23

Disk Access Time = seek time + latency

Ave. Seek Time

http://www.storagenewsletter.com/news/disk/hdd-technology-trends-ibm

Tom Coughlan24

Disk Latency

7,200 RPM

10,000 RPM

15,000 RPM

http://www.storagenewsletter.com/news/disk/hdd-technology-trends-ibm

Tom Coughlan25

Conclusions (cont.)

● Flash has arrived just in time.

● Shared storage (NAS, SAN) will remain prominent, with additional emphasis on cost effective scale-out.

● Gluster● pNFS● iSCSI, FCoE

● Initiator and target

● More storage boxes => better management is required:● libStorageMgmt

● http://sourceforge.net/projects/libstoragemgmt/

● An opportunity for Linux as a low-cost scale-out storage server.

Tom Coughlan26

Thank-you.