diablo technologies - flash memory summit technologies slide 2 $ dram capacity is limited while cost...

Diablo Technologies Slide 1

Diablo Technologies

Memory1™

Maher Amer

CTO Diablo Technologies - Highly Confidential


$DRAM capacity is limited

while cost is prohibitive

THE NEED: Big Data applications need Big Memory

THE SOLUTION: Expand system memory with Memory1

Big Memory for Big Data


Two Tier Memory Sub-Systems are Here to Stay

Load

/Sto

re(A

pp

licat

ion

Dir

ect

Acc

ess)

Block(Application Indirect Access)

CPU

Registers

Cache

DRAM

Storage

Byte-addressable with Flash Capacity and Economics


Memory1 At A Glance

TM

• DDR4 memory DIMM with up to 128 GB per module

• Economically expands application memory by terabytes

• No changes required to servers, applications

• Industry standard, JEDEC-compliant LRDIMM/RDIMM


How It Works: High-Performance Hardware Solution

Software Intelligently Manages Memory Access

• Acts as extension of OS Virtual Memory Manager

• Implements intelligent paging algorithms

• Optimizes performance and extends Flash endurance

Modules Plug Into DDR4 Memory Slots

• Memory bus is highest performing interface to CPU

• Over 17GB/s per memory channel @ 2133 MT/s

• Lower latency than PCIe/NVMe

Innovative All-Flash DDR4 DIMM Hardware

• Deployed in parallel with standard DRAM

• Leverages Flash capacity, power, and cost advantages

QP

I Memory BusQPI


Target Applications and Workloads

BIG DATA

PROCESSING

Real-time analysis

Distributed caching

Predictive analytics

Caching

Paid search

Key-value lookup

CLOUD

Distributed database

In-memory database

Relational database

DATABASE

Application data doesn’t fit in one

machine


machine but DRAM constrained


machine


Diablo Memory Expansion


Hardware

Combination of firmware and software• Intelligently manages application memory

access

• Leverages CPU hardware and special statistics

• Manages performance and endurance

No Application Changes• Loads like a driver

Two Major Components• Data Management

• Media Management

Diablo Memory Expansion (DMX) Software

Application

Operating System

DMX Software

RAM


DMX Software Intelligently Manages Expanded Memory

DATA MANAGEMENTKeeps track of all application page sequences and

ensures correct pages are in DRAM

MEDIA MANAGEMENT Organizes and optimizes the flash layer for maximum

performance and endurance

• DMX Software loads and operates as an OS-level driver

• Application Software requires no changes

APPLICATION

SOFTWARE

DRAM

Memory1 Flash Memory (Total Available System Memory)


Data Management Details

Data Tiering• DMX keeps hot data in DRAM

• High priority data maintained in DRAM

• Cold data evicted to flash

Quality of Service• Priority Associated Data Placement

• Additional DRAM allocated per

application increasing hot data in DRAM

• Keeps data with response time

requirements in DRAM

Data

Tiering

Quality

of

Service


Data Management Details

CPU

CPU

NUMA Node

Amortized PF

Data Pages

Clustering

Smart

Prefetch

DRAM

DRAM

Learning Engine

• Application Profiling and Analytics

- Monitors application data access behaviors

• Data access prediction

- Predicts next or additional pages required

• Smart Data pre-fetch

- Pre-fetches pages to DRAM based on profiling,

history, and data access patterns

Clustered Pages• Prefetches grouped data typical in many applications

Data Locality• Movement between DRAM and Memory1 ensures data

local to associated node

Amortized Page Faults• Groups page requests together, fully leveraging page fault


Media Management Details

Flash Management

• Low-level media management (handled in

firmware)

• SoftFTL – Adaptation for 4K pages

• Tuneable cache ratio (4:1, 8:1)

• Device Striping

Intelligent Traffic Management

• Dirty Page Writes

- Avoids premature writes to flash for frequently

written pages

- Minimizes Read/Modify/Write operations

• Traffic Sequentialization

- Pages evicted are written sequentially to flash

DRAM

Evictions

Written

Sequentially

Clean

Pages

Dropped

Dirty Pages

Remain in

DRAM Until

Writes

Complete

Pages

Naturally

Pulled in by

Application


Application Benchmarks


Kdb+ Time Series Software: Memory1 Use Case

Generate Stock

Data

Generate

Returns

Generate

Moving

Averages

Generate

Statistics

Stock Ticker Analysis and Regression:

300GB of in-memory data generated (after Garbage Collection)

2HRs of Execution Time on 2S EP Platform w/ 512GB DRAM

Kdb+ time series data software | Kx SystemsThe world’s most powerful number cruncher, kdb+ offers unparalleled

performance for time-series data and analytics

https://kx.com/software.php


Kdb+: 2TB Memory1 Configuration

Memory1 2TB KDB+ Appliance

CPU: Haswell/Broadwell

Cores: 14-18C/socket

DRAM: 256GB

MEMORY1: 1900GB

NIC: built-in GE + Add-on 10GE

Storage: N x HDD Increase dataset size and instances per

machine by 5X and avoid EX Platforms

0 0.5 1 1.5 2 2.5 3 3.5

DRAM Instance

M1 Instance1

M1 Instance2

M1 Instance3

M1 Instance4

M1 Instance5

Execution Time (Hrs)


Memory1 Advantage: Economically Increases MySQL Buffer Pool

RAM MEMORY1

xMB yGB

zGB MySQL Instance

Buffer Pool

Instance NMain and

Virtual Memory

(includes all

buffers)

Physical Disk/Secondary

Storage (log files, databases

and relative statistics, index,

data, and meta-data files

Buffer

Pools

MySQL Architecture

STORAGE CAPACITY

Significantly Expand Cache

Increase TPS

Improve Response Time


Test Configuration

Random, Normal Workload Distribution

“DRAM-Only” Server 10GE Optical NIC

Dual 14-Core CPUs

128GB DRAM

Centos7

MySQL:

Memory1 Server10GE Optical NIC

Dual 14-Core CPUs

128GB DRAM

1TB Memory1

Centos7

MySQL:

Client (Load Generator)10GE Optical NIC

Linux:

Sysbench:

RAM MEMORY1

STORAGE CAPACITY

Increasing Buffer Pool size

with MEMORY1 should:

Remove Bottleneck imposed

by Storage Performance

Increase TPS and Reduce

Response Time

10GE Optical Switch


Hardware Configuration

DRAM / NVMe ServerDual 14-Core CPUs

128GB DRAM

NVMe SSD as storage

Memory1 ServerDual 14-Core CPUs

128GB DRAM

1TB Memory1

Client (Load Generator)

Sysbench

Increasing Buffer Pool size

with Memory1 will:

• Remove bottleneck imposed by

storage performance

• Increase TPS and reduce response

time

• Allow top performance with any

storage solution

All Servers Include: 10GE Optical NIC CentOS 7

10GE Optical Switch

Random, Normal Workload Distribution

RAM MEMORY1

STORAGE CAPACITY


Benchmark Scenarios

75

GB

DB

80

GB

Bu

ffer-P

oo

l

75

GB

DB

15 G

B

Bu

ffer-

Po

ol

Scenario “A” Scenario “B”

DMX:

6x Scenario A

Total RAM used = 128G

Sysbench Mixed Traffic

RAM:

6x Scenario B

Total RAM used = 128G

Sysbench Mixed Traffic

15 G

B

Ca

ch

e


DMX vs RAM

0

200

400

600

800

1000

1200

1400

1600

1800

13

51

70

11

05

11

40

11

75

12

10

12

45

12

80

13

15

13

50

13

85

14

20

14

55

14

90

15

25

15

60

15

95

16

30

16

65

1

TPS

Time (sec)

TPS Comparison

RAM

DMX

0

10

20

30

40

50

60

70

80

1

38

6

77

1

11

56

15

41

19

26

23

11

26

96

30

81

34

66

38

51

42

36

46

21

50

06

53

91

57

76

61

61

65

46

Late

ncy

(m

s)

Time (Sec)

Latency Comparison

RAM Latency

DMX Latency


Storage Read BW


Maher Amer

CTO, Diablo Technologies

[email protected]

Thank You!

diablo technologies - flash memory summit technologies slide 2 $ dram capacity is limited while cost...

Documents