8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
1/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Shifting Gearswith SSDsSNIA Summer Symposium
Denis Vilfort
2008 Sun Microsystems, Inc.1
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
2/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
What Well CoverIn the next 45 minutes
Why SSDs Now?
Were do SSDs Fit?
SSDs Impact onNetworked Storage
New ArchitectureDecisions with SSD
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
3/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Web 2.0The Long Tail of Data
Online services carry far moreinventory than traditional retailers
Example: Rhapsody caries 56
times as many songs as Wal-Marts 36,000.
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
4/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Global Data ContinuumLatency Differentiated Storage Pools
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
5/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Where to Store Data?Optimization Trade-Off
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
6/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Disk to CPU DiscontinuityTodays CPU have become I/O starved
Moores Law is out-stripping disk drive
rotational speed As a result, systems are
hopelessly unbalanced
between CPU capabilityand storage poolperformance
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
7/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Enterprise HDD Falling ShortIncreasing Capacity to Performance Mismatch
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
8/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Storage EvolutionTime for a Change?
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
9/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Why Applications Dont PerformWaiting for DATA
TodaysMulti-Core, Multi-Socketapplication server design areincreasingly held back by slowstorage
When requesting data, the serverspendsmost of its time waitingfor storage
Application performanceremainsluggishregardless of the ServerCPU horsepower
The traditional remedy of addingmore expensiveDRAM may nolonger sufficeas data sets doubleevery 2 years
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
10/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Turbo Charged ApplicationsSSDs Eliminate Storage Bottlenecks
TodaysMulti-Core, Multi-Socket
application server design are nowserved by High PerformanceSSDs
The serverno longer waists time
waiting for data Applicationperformance isas
highas the Server CPUshorsepower
No longer need to remedysluggish storage performance byadding expensiveDRAM
Bottom Line:Improved end user experience,faster results & Better ROI
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
11/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
CPU to Storage DiscontinuityThe number of HDDs needed to keep up
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
12/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
SSDs Keep Up w/ CPUMoores Law controls both
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
13/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
New Server Memory Hierarchy
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
14/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Latency ComparisonBridging the DRAM to HDD Gap
1nS
10nS
100nS
1uS
10uS
100uS
1mS
10mS
100mS
1 S
TAPE
HDDFLASH/
SSD
DRAM
CPU
100,000 X
Latency
Mismatch
150 X
LatencyReduction
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
15/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Solid State Drives (SSD)Enterprise advantage from commodity FLASH SSD has three major parts:
> A) Controller
> B) DRAM
> C) FLASH bank
Individual FLASH chips arepooled andaddress space virtualized by thecontroller
Controller also performs
> Wear leveling
> CRC
> Bad block mapping
Controller provides the host interfacesuch as SATA, PATA, SAS or FC
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
16/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Anatomy of an SSDSimilar to HDD Arrays
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
17/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
SSDs Why Now?$/GB Crossover in 2009
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
18/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
SSD to HDD ComparisonSSD for Performance, HDD for Capacity
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
19/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Where to Deploy SSDs?Storage or Server?
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
20/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
IT SSD Strategy #1HDD Replacement
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
21/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Modern CPUs are capable of
large amounts of IO HDDs are great capacity devices
- but poor IO devices
SSDs are great IO devices- but costly capacity devices
Balanced System Design> CPU IO = Storage IO
> Capacity = Data Set
Sun Optimized System DesignBest Application Performance
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
22/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Classic CPU-Storage ImbalanceSunFire x4450 Memory Hierarchy
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
23/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
The Hybrid Storage PoolGaining Performance and Capacity Balance
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
24/32SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Building the Hybrid Storage PoolCombining Performance and Capacity
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
25/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
ZFS Turbo Charges ApplicationsThe Hybrid Storage Pool
ZFS automatically :>Writes new data to a very fast SSD
pool (ZIL)
>Determines data access patterns
andstores frequently accessed
data in the L2ARC
>Bundles IO into sequential lazywrites for moreefficient use of low
cost mechanical disks
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
26/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Hybrid Storage Pool EconomicsRight-Size Performance & Capacity
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
27/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
Standard HDDs Starve ServersCache vs. SATA Storage Pool
Point to Point Switched Backplane
200,000 IOPS
20%
80%
40,000
IOPS
2,688IOPS
42,688IOPS100%
48 X 70 IOPS = 3,360 IOPS
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
28/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
SSDs Turbo Charges ServersCache vs. Hybrid Storage Pool
Point to Point Switched Backplane
200,000 IOPS
20%
80%
40,000
IOPS
160,000IOPS
200,000IOPS100%
6 X 35,000 IOPS = 210,000 IOPS 42 X 70 IOPS = 2,940 IOPS
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
29/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
High Performance Storage ServerParallel Hybrid Storage Pools
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
30/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
High Performance Storage ServerSun Fire X4540 Example
HBA
105K IOPSSSDs
70K IOPS
CPU Complex
380K IOPS
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
31/32
SNIA Summer Symposium 2008 Sun Microsystems, Inc.
In Summary
FLASH iscommoditizingdriven by ConsumersElectronics
SSDs will beEverywhere
including the Data CenterwherePowerandPerformanceare concerns
Hybrid Storage Poolsarenow the best strategy and Sun is leading the way
withZFS
8/6/2019 Denis Vilfort_Sun_Shifting Gears With SSDs
32/32
Shifting Gears with SSDs
Thank You
Denis Vilfort
[email protected] (408) 896 4290
2008 Sun Microsystems, Inc.
mailto:[email protected]:[email protected]