A State Machine Architecture for High- Efficiency, Low ... · PDF file• Want SSD Performance & Latency ... Tradeoffs between Cost, Performance and Life ... Ground Up Design...

A State Machine Architecture for High-Efficiency, Low-Latency SS Media

Performance Extraction

Bret S. Weber DataDirect Networks

Flash Memory Summit 2012 Santa Clara, CA

1

The Real Issue

• Want SSD Performance & Latency • Need Enterprise Redundancy • Use Commodity Flash Products

• Non Proprietary • Customer Replaceable • No Technology “Lock In” • Lowest $/TB

• Allow Seamless Levels of Storage • Low $/TB • Low $/IOP

2

SSD Small Block IOPs

3

SSD Large Sequential

4

Flash Performance Takeaways

► Flash Performance is All over the Map ► The Technology Changes Fast ► Tradeoffs between Cost, Performance and Life ► Investment Protection

5

Typical Controller IOPs Throttling

6

0

200,000

400,000

600,000

800,000

1,000,000

1,200,000

1,400,000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Small Block IOPs

SSD Capability Typical Storage Controller Capability

Typical Controller BW Throttling

7

0

2000

4000

6000

8000

10000

12000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Bandwidth (MB/S)

SSD Capability Typical Storage Controller Capability

Storage Fusion Architecture History

► Previous Experience with Silicon-Based RAID-3 Arrays • High Bandwidth • Very consistent Quality of Service (Bandwidth and Latency)

► Ground Up “Blank Sheet of Paper” Architecture in 2007 • Typical Five Year Maturity Cycle

► Address Emerging Applications • HPC • Data-Intensive Cloud Applications • Big Data Analytics

► Address the Emerging Technologies • Multi-Core Processors • High IOP Non Volatile Memory Technologies • Server Virtualization • Scale Out

Ground-Up Design Big Data Optimized

SFA – Details

► Ground Up Design Exploiting Emerging Technologies ► Optimized for Ultra-High IOPs & Bandwidth

• Application Space Code • Kernel Bypass for all IO Operations • Real-Time NUMA Scheduler Minimizes Latency • Built to Exploit All Available CPU Cores • BW, IOPs and Mixed Workload Configurations

► Fully Parallel I/O Execution Engine • No Lock Architecture

► Portable Code • Rapid Time To Market • Optimizations around Linux architectures

► Highest levels of Performance & QoS ► DDN device drivers accelerate I/O ► Optional Server Virtualization Brings Big Data Closest to Processing

240Gb/s Cache Link

32-64GB High-Speed Cache

32-64GB High-Speed Cache

SFA Interface Virtualization SFA Interface Virtualization

960Gb/s Internal SAS Storage Management Network

16 x FDR InfiniBand Host Ports

SFA RAID 5,6

RAID 5,6

SFA RAID 1

1 2 3 4 5 6 7 8 P

RAID 5,6

Q RAID 6

1 2 3 4

1 1m

Q RAID 6

P RAID 5,6

Internal SAS Switching

Internal SAS Switching

Typical Approaches

10

User Application

GNU C Library

System Call Interface

Kernel-Level Engine

Architecture Dependent Kernel Code

User Space

OS Kernel

Kernel Level Implementation

Data Path

Sys

tem

Man

agem

ent

Data Path

Storage Application

GNU C Library

System Call Interface

Kernel

Architecture Dependent Kernel Code

User Space

OS Kernel

Typical User Space Implementation

Sys

tem

Man

agem

ent

Kernel Context Switches Add Significant Latency

In-Kernel Dependencies Prohibit Easy Portability

Low Latency – No Context Switching

User Application

GNU C Library

System Call Interface

Kernel-Level Engine

Architecture Dependent Kernel Code

User Space

OS Kernel

Kernel Level Implementation

Data Path

Sys

tem

Man

agem

ent

Data Path

DDN SFA Application

GNU C Library

System Call Interface

Kernel

Architecture Dependent Kernel Code

User Space

OS Kernel

DDN User Space Implementation

Sys

tem

Man

agem

ent

Data Path

Storage Application

GNU C Library

System Call Interface

Kernel

Architecture Dependent Kernel Code

User Space

OS Kernel

Typical User Space Implementation

Sys

tem

Man

agem

ent

► No Context Switching ► Low Latency IO ► Predictable Performance ► High Speed Routing

Architecture

A Scalable & Abstractable I/O Delivery Architecture

12

Front-End Services Abstraction

Back-End Services Abstraction

SFA JIPC: Non-Preemptive

I/O Routing System

InfiniBand RDMA Driver Fibre Channel Driver Memory DMA Driver

Internal VM Clients External Block-Based Applications

User Space I/O Scheduler

Owns The Kernel & Prohibits Disruption

(State Machine)

I/O Routing, RAID Services, Cache Management & Disk Virtualization

Com

plet

e St

orag

e O

S

Scale Scale

SFA12K Architecture

► 40GB/S Throughput ► 1.4M IOPs to SSD ► DirectFlash Caching

Acceleration

State Machine

Scale Up To Any # Of CPU Cores Today: 8/Socket

Scale Out To Any # Of CPU Sockets

Today: 4 Sockets

SFA Controller IOPs – No Throttling

14

0

200,000

400,000

600,000

800,000

1,000,000

1,200,000

1,400,000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Small Block IOPs

SSD Capability Typical Storage Controller Capability

SFA with SSD

SFA Linear Scaling (SFA10K)

15

0

100000

200000

300000

400000

500000

600000

700000

800000

1 4 8 12 16 20 24 28 32 36

IO/S

Pool Count

random aligned read, Rate (IO/s), Varying Pool Count

SFA10K - ssd_1+1_WM_Re,4K, QD 16

Previous Generation Product

SFA Controller – No BW Throttling

16

0

2000

4000

6000

8000

10000

12000

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Bandwidth (MB/S)

SSD Capability Typical Storage Controller Capability

SFA with SSD

SFA Linear Scaling (SFA10K)

17

0

2000

4000

6000

8000

10000

12000

14000

1 4 8 12 16 20 24 28 32 36 44 48 50 56

MB

/S

Pool Count

random aligned read, Rate (IO/s), Varying Pool Count

SFA 10K - ssd_1+1_WM_Re,64K, QD 16

Previous Generation Product

Our Surprise: 4KB I/O (SFA10K) Latency That Defies Convention

1 x 8+2: Pliant 800GB 1 x 4+1: Pliant 800GB QD Avg Latency (μs) Avg Latency (μs)

1 75.87 79.15 2 62.77 69.78 4 72.58 76.89 8 78.21 84.97

16 86.80 90.98 32 95.67 96.79 64 102.17 107.89

128 99.88 105.70

18

1 x 8+2: Pliant 800GB 1 x 4+1: Pliant 800GB QD Avg Latency (μs) Avg Latency (μs)

1 324.78 323.02 2 156.92 156.07 4 72.91 73.11 8 37.86 38.15

16 19.97 20.34 32 10.38 11.16 64 6.51 7.58

128 6.89 8.23

Ran

dom

Writ

e R

ando

m R

ead

Actual Case Studies

► Rich Imagery Processing Application

► GPFS File System ► Mix of Small & Large File

Requests (mostly small) ► SSD System Bake-Off ► DDN Selected Due To:

• Multi-Dimensional Performance • Low Latency & Wall Clock • Lowest Data Center Footprint

19

Leading U.S. Cloud-Based Applications

Provider

SFA10K-M Systems QDR InfiniBand Connected

114 x 800GB SSDs

Conclusions

► Get Performance and Redundancy ► SSDs Can Compete With The Right Architecture ► Lots Of Options: Don’t Lock Yourself In ► Don’t artificially limit your performance ► Big Data Requires All Varieties of Performance

• We don’t know what we don’t know…

7/27/2012 20

Questions ???

7/27/2012 21