Bringing Your Applications to the Fast Lane
Deepak Singh - Amazon Web Services
Christos Kalantzis - Netflix
Steven Jones -...
Hello!
Friday, November 15, 13
3
Friday, November 15, 13
Instance Types

Friday, November 15, 13
Performance

Friday, November 15, 13
Generations

Friday, November 15, 13
Many instance types

Friday, November 15, 13
Friday, November 15, 13
Building Blocks

Friday, November 15, 13
General Purpose
Compute-optimized
Memory-optimized
Storage-optimized
GPU
Micro

Friday, November 15, 13

M1, M3
C1, CC2
M2...
Workload optimized

Friday, November 15, 13
Compute-optimized

Friday, November 15, 13
CC2

Friday, November 15, 13
2.6 GHz Intel Xeon E5-2670

Friday, November 15, 13
16 cores/32 vCPU

Friday, November 15, 13
10 Gbps cluster networking

Friday, November 15, 13
High Performance Linpack

Friday, November 15, 13
Cores

Friday, November 15, 13

max
R

240.09

peak
R

cc2.8xlarge

17024

354.12
I/O

Friday, November 15, 13
HI1 instances

Friday, November 15, 13
High random I/O performance

Friday, November 15, 13
SSDs

Friday, November 15, 13
Netflix and HI1
Christos Kalantzis

Friday, November 15, 13
Advantages of using hi1.4xlarge with C*
• Higher application speed
• Efficient operations & better data quality
• Cost sav...
Higher Application Speeds
• Compared m2.4xlarge & hi1.xlarge
• 6 node C* cluster with 250 GB data on each node
• Used foll...
Higher Application Speeds

C*

Cassandra-stress

Cassandra-stress

Cassandra-stress

Friday, November 15, 13

C*

C*

C*
C...
m2.4xlarge results
Throughput

Friday, November 15, 13

Speed
hi1.4xlarge results
Throughput

Friday, November 15, 13

Speed
Higher Application Speeds
• 40X throughput
• Better Latencies
– ~37X AVG
– ~109X 95th %ile
– ~110X 99th %ile

Friday, Nove...
Efficient C* Operations & Better Data Quality
• Cassandra => “benefit now, pay later”
– Immutable SSTABLES need to be comp...
Cost Savings
• Can now have multi-tenant clusters
– Savings immediate on the second application
– m2.4xlarge is $1.640 per...
Cost Savings
• Use Less Nodes
– Data : Memory ratio can now be bigger
– Cluster size doesn’t need to be doubled as often

...
Cost Savings
• Smaller Operational Footprint
– Less clusters
– Smaller operational teams
– Less fire-fighting
– Team can f...
hi1.4xlarge caveats
• 10 Gbps not attained (in our model)
– Need to be part of the same Cluster Placement Group
– Netflix ...
Memory-optimized

Friday, November 15, 13
244 GiB of RAM

Friday, November 15, 13
2.6 GHz Intel Xeon E5 2670

Friday, November 15, 13
10 Gbps cluster networking

Friday, November 15, 13
In-memory distributed analytics

Friday, November 15, 13
e.g. SAP HANA

Friday, November 15, 13
Steven Jones
Amazon Web Services

Friday, November 15, 13
SAP

Friday, November 15, 13

ERP
SAP

Friday, November 15, 13

ERP
SAP

Friday, November 15, 13

ERP
SAP

Friday, November 15, 13

ERP

BI
SAP

Friday, November 15, 13

ERP

BI
SAP

Friday, November 15, 13

ERP

BI
SAP

Friday, November 15, 13

ERP

BI
Aggregates

SAP

Friday, November 15, 13

ERP

BI
Aggregates

SAP

Friday, November 15, 13

ERP

Pre-Defined
Queries

BI
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Pre-Defined
Queries

Aggregates

SAP

Friday, November 15, 13

ERP

BI

BWA

TREX
Magnetic Disk
SSD/Flash
Latency
RAM
CPU Cache
CPU Register
Storage Type
Friday, November 15, 13
Magnetic Disk
RAM is up to 100,000
x Faster

SSD/Flash
Latency
RAM

CPU Cache
CPU Register
Storage Type
Friday, November 1...
SAP HANA
High Performance Analytic Appliance
aka In-Memory Database Platform

http://tinyurl.com/hana-perf
Friday, Novembe...
SAP

Friday, November 15, 13

ERP

BI
HANA
SAP

Friday, November 15, 13

ERP
HANA
EC2 Cluster Compute instances
cc2.8xlarge

cr1.8xlarge

2 x Intel Xeon E5-2670 processors
32 vCPUs with hyperthreading
64-...
SAP HANA on AWS
1,776 Core HANA Cluster on AWS
60B Rows | 8M Rows/s | 3M Queries/h | 330ms

Friday, November 15, 13
Where to go for HANA
SAP Developer Center
http://bit.ly/aws-hana-dev
Free SAP license

AWS Marketplacehttp://bit.ly/
aws-h...
New generations

Friday, November 15, 13
Latest Generation
Hardware

Friday, November 15, 13
Hardware features,
e.g. AVX

Friday, November 15, 13
SSDs

Friday, November 15, 13
Enhanced Networking

Friday, November 15, 13
Better priceperformance

Friday, November 15, 13
C3

Friday, November 15, 13
Best raw compute capacity

Friday, November 15, 13
2.8 GHz Intel Xeon E5-2680v2
(Ivy Bridge)

Friday, November 15, 13
Clustering as a “feature”

Friday, November 15, 13
Friday, November 15, 13
Friday, November 15, 13
High Performance Linpack

Friday, November 15, 13
Cores

Friday, November 15, 13

max
R

484.18

peak
R

c3.8xlarge

26496

593.87
Rmax%
600"

500"

400"

300"

200"

100"

0"
CC1"

Friday, November 15, 13

CC2"

C3"
Cores

Friday, November 15, 13

max
R

163.9

peak
R

c3.8xlarge

8192

183.5
I2

Friday, November 15, 13
Coming Soon

Friday, November 15, 13
Best Price for IOPS

Friday, November 15, 13
More sizes

Friday, November 15, 13
More SSD

Friday, November 15, 13
More IOPS

Friday, November 15, 13
More memory

Friday, November 15, 13
Friday, November 15, 13
350,000+ random read IOPS
320,000+ random write IOPS
i2.8xlarge

Friday, November 15, 13
175,000+ random read IOPS
160,000+ random write IOPS
i2.4xlarge

Friday, November 15, 13
Enhanced Networking

Friday, November 15, 13
Enhanced Networking
Low latency
Low jitter
Very high PPS performance

Friday, November 15, 13
Latency

cc2.8xlarge+

c3.8xlarge+

c3.8xlarge+(enhanced)+

PPS

cc2.8xlarge+

Friday, November 15, 13

c3.8xlarge+

c3.8x...
p50$
p90$
p999$
p100$

Low network jitter

Friday, November 15, 13
General Purpose
Compute-optimized
Memory-optimized
Storage-optimized
GPU
Micro

Friday, November 15, 13

M1, M3
C1, CC2, C...
General Purpose
Compute-optimized
Memory-optimized
Storage-optimized
GPU
Micro

Friday, November 15, 13

M1, M3
C1, CC2, C...
More to come

Friday, November 15, 13
Please give us your feedback on this
presentation

CPN203
As a thank you, we will select prize
winners daily for completed...
Upcoming SlideShare
Loading in …5
×

Bringing Your Applications to the Fast Lane (CPN203) | AWS re:Invent 2013

831 views

Published on

Amazon Elastic Compute Cloud (Amazon EC2) has added a number of instance types that provide a high level of performance. Instances range from compute-optimized instances to instances that deliver thousands of IOPS. In this session, you will learn more about Amazon EC2 high performance instance types and hear from customers about how they are using these instances to improve application performance, and reduce costs.

Published in: Technology
0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
831
On SlideShare
0
From Embeds
0
Number of Embeds
3
Actions
Shares
0
Downloads
23
Comments
0
Likes
1
Embeds 0
No embeds

No notes for slide

Bringing Your Applications to the Fast Lane (CPN203) | AWS re:Invent 2013

  1. 1. Bringing Your Applications to the Fast Lane Deepak Singh - Amazon Web Services Christos Kalantzis - Netflix Steven Jones - Amazon Web Services November 15, 2013 © 2013 Amazon.com, Inc. and its affiliates. All rights reserved. May not be copied, modified, or distributed in whole or in part without the express consent of Amazon.com, Inc. Friday, November 15, 13
  2. 2. Hello! Friday, November 15, 13
  3. 3. 3 Friday, November 15, 13
  4. 4. Instance Types Friday, November 15, 13
  5. 5. Performance Friday, November 15, 13
  6. 6. Generations Friday, November 15, 13
  7. 7. Many instance types Friday, November 15, 13
  8. 8. Friday, November 15, 13
  9. 9. Building Blocks Friday, November 15, 13
  10. 10. General Purpose Compute-optimized Memory-optimized Storage-optimized GPU Micro Friday, November 15, 13 M1, M3 C1, CC2 M2, CR1 HI1, HS1 CG1, G2 Micro
  11. 11. Workload optimized Friday, November 15, 13
  12. 12. Compute-optimized Friday, November 15, 13
  13. 13. CC2 Friday, November 15, 13
  14. 14. 2.6 GHz Intel Xeon E5-2670 Friday, November 15, 13
  15. 15. 16 cores/32 vCPU Friday, November 15, 13
  16. 16. 10 Gbps cluster networking Friday, November 15, 13
  17. 17. High Performance Linpack Friday, November 15, 13
  18. 18. Cores Friday, November 15, 13 max R 240.09 peak R cc2.8xlarge 17024 354.12
  19. 19. I/O Friday, November 15, 13
  20. 20. HI1 instances Friday, November 15, 13
  21. 21. High random I/O performance Friday, November 15, 13
  22. 22. SSDs Friday, November 15, 13
  23. 23. Netflix and HI1 Christos Kalantzis Friday, November 15, 13
  24. 24. Advantages of using hi1.4xlarge with C* • Higher application speed • Efficient operations & better data quality • Cost savings Friday, November 15, 13
  25. 25. Higher Application Speeds • Compared m2.4xlarge & hi1.xlarge • 6 node C* cluster with 250 GB data on each node • Used following Cassandra Stress tool & Options from three m2.4xlarge clients: – cassandra-stress -d [server-list] -t 25 -r -p 7102 -n 1000000000 -k -o READ Friday, November 15, 13
  26. 26. Higher Application Speeds C* Cassandra-stress Cassandra-stress Cassandra-stress Friday, November 15, 13 C* C* C* C* C*
  27. 27. m2.4xlarge results Throughput Friday, November 15, 13 Speed
  28. 28. hi1.4xlarge results Throughput Friday, November 15, 13 Speed
  29. 29. Higher Application Speeds • 40X throughput • Better Latencies – ~37X AVG – ~109X 95th %ile – ~110X 99th %ile Friday, November 15, 13
  30. 30. Efficient C* Operations & Better Data Quality • Cassandra => “benefit now, pay later” – Immutable SSTABLES need to be compacted – Can compact faster & with less impact • Eventual Consistency – Repairs run more often – Increased Consistency without higher CL calls Friday, November 15, 13
  31. 31. Cost Savings • Can now have multi-tenant clusters – Savings immediate on the second application – m2.4xlarge is $1.640 per hour – hi1.4xlarge is $3.100 per hour C* App A C* App A C* App A C* App A C* App A Friday, November 15, 13 C* App B C* App A C* App A&B C* App B C* App B C* App B C* App B C* App B C* App A&B C* App A&B C* App A&B C* App A&B C* App A&B
  32. 32. Cost Savings • Use Less Nodes – Data : Memory ratio can now be bigger – Cluster size doesn’t need to be doubled as often Friday, November 15, 13
  33. 33. Cost Savings • Smaller Operational Footprint – Less clusters – Smaller operational teams – Less fire-fighting – Team can focus on automation Friday, November 15, 13
  34. 34. hi1.4xlarge caveats • 10 Gbps not attained (in our model) – Need to be part of the same Cluster Placement Group – Netflix cross-zone availability model excludes CPG • Not same as bare metal SSD running on Linux – Virtualization obfuscates some SSD advantages Friday, November 15, 13
  35. 35. Memory-optimized Friday, November 15, 13
  36. 36. 244 GiB of RAM Friday, November 15, 13
  37. 37. 2.6 GHz Intel Xeon E5 2670 Friday, November 15, 13
  38. 38. 10 Gbps cluster networking Friday, November 15, 13
  39. 39. In-memory distributed analytics Friday, November 15, 13
  40. 40. e.g. SAP HANA Friday, November 15, 13
  41. 41. Steven Jones Amazon Web Services Friday, November 15, 13
  42. 42. SAP Friday, November 15, 13 ERP
  43. 43. SAP Friday, November 15, 13 ERP
  44. 44. SAP Friday, November 15, 13 ERP
  45. 45. SAP Friday, November 15, 13 ERP BI
  46. 46. SAP Friday, November 15, 13 ERP BI
  47. 47. SAP Friday, November 15, 13 ERP BI
  48. 48. SAP Friday, November 15, 13 ERP BI
  49. 49. Aggregates SAP Friday, November 15, 13 ERP BI
  50. 50. Aggregates SAP Friday, November 15, 13 ERP Pre-Defined Queries BI
  51. 51. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  52. 52. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  53. 53. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  54. 54. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  55. 55. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  56. 56. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  57. 57. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  58. 58. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  59. 59. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  60. 60. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  61. 61. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  62. 62. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  63. 63. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  64. 64. Pre-Defined Queries Aggregates SAP Friday, November 15, 13 ERP BI BWA TREX
  65. 65. Magnetic Disk SSD/Flash Latency RAM CPU Cache CPU Register Storage Type Friday, November 15, 13
  66. 66. Magnetic Disk RAM is up to 100,000 x Faster SSD/Flash Latency RAM CPU Cache CPU Register Storage Type Friday, November 15, 13
  67. 67. SAP HANA High Performance Analytic Appliance aka In-Memory Database Platform http://tinyurl.com/hana-perf Friday, November 15, 13
  68. 68. SAP Friday, November 15, 13 ERP BI HANA
  69. 69. SAP Friday, November 15, 13 ERP HANA
  70. 70. EC2 Cluster Compute instances cc2.8xlarge cr1.8xlarge 2 x Intel Xeon E5-2670 processors 32 vCPUs with hyperthreading 64-bit 60.5 GB RAM 10 Gigabit Network 2 x Intel Xeon E5-2670 processors 32 vCPUs with hyperthreading 64-bit 244 GB RAM 10 Gigabit Network NUMA and Turbo Support New All c* instances support high-performance (10 gigabit) networking to Elastic Block Storage volumes (EBS) Friday, November 15, 13
  71. 71. SAP HANA on AWS 1,776 Core HANA Cluster on AWS 60B Rows | 8M Rows/s | 3M Queries/h | 330ms Friday, November 15, 13
  72. 72. Where to go for HANA SAP Developer Center http://bit.ly/aws-hana-dev Free SAP license AWS Marketplacehttp://bit.ly/ aws-hana-one 99¢ / hr SAP License SAP HANA Marketplace http://bit.ly/aws-hana-byol BYOL Friday, November 15, 13
  73. 73. New generations Friday, November 15, 13
  74. 74. Latest Generation Hardware Friday, November 15, 13
  75. 75. Hardware features, e.g. AVX Friday, November 15, 13
  76. 76. SSDs Friday, November 15, 13
  77. 77. Enhanced Networking Friday, November 15, 13
  78. 78. Better priceperformance Friday, November 15, 13
  79. 79. C3 Friday, November 15, 13
  80. 80. Best raw compute capacity Friday, November 15, 13
  81. 81. 2.8 GHz Intel Xeon E5-2680v2 (Ivy Bridge) Friday, November 15, 13
  82. 82. Clustering as a “feature” Friday, November 15, 13
  83. 83. Friday, November 15, 13
  84. 84. Friday, November 15, 13
  85. 85. High Performance Linpack Friday, November 15, 13
  86. 86. Cores Friday, November 15, 13 max R 484.18 peak R c3.8xlarge 26496 593.87
  87. 87. Rmax% 600" 500" 400" 300" 200" 100" 0" CC1" Friday, November 15, 13 CC2" C3"
  88. 88. Cores Friday, November 15, 13 max R 163.9 peak R c3.8xlarge 8192 183.5
  89. 89. I2 Friday, November 15, 13
  90. 90. Coming Soon Friday, November 15, 13
  91. 91. Best Price for IOPS Friday, November 15, 13
  92. 92. More sizes Friday, November 15, 13
  93. 93. More SSD Friday, November 15, 13
  94. 94. More IOPS Friday, November 15, 13
  95. 95. More memory Friday, November 15, 13
  96. 96. Friday, November 15, 13
  97. 97. 350,000+ random read IOPS 320,000+ random write IOPS i2.8xlarge Friday, November 15, 13
  98. 98. 175,000+ random read IOPS 160,000+ random write IOPS i2.4xlarge Friday, November 15, 13
  99. 99. Enhanced Networking Friday, November 15, 13
  100. 100. Enhanced Networking Low latency Low jitter Very high PPS performance Friday, November 15, 13
  101. 101. Latency cc2.8xlarge+ c3.8xlarge+ c3.8xlarge+(enhanced)+ PPS cc2.8xlarge+ Friday, November 15, 13 c3.8xlarge+ c3.8xlarge+(enhanced)+
  102. 102. p50$ p90$ p999$ p100$ Low network jitter Friday, November 15, 13
  103. 103. General Purpose Compute-optimized Memory-optimized Storage-optimized GPU Micro Friday, November 15, 13 M1, M3 C1, CC2, C3 M2, CR1 HI1, I2, HS1 CG1, G2 Micro
  104. 104. General Purpose Compute-optimized Memory-optimized Storage-optimized GPU Micro Friday, November 15, 13 M1, M3 C1, CC2, C3 M2, CR1 HI1, I2, HS1 CG1, G2 Micro
  105. 105. More to come Friday, November 15, 13
  106. 106. Please give us your feedback on this presentation CPN203 As a thank you, we will select prize winners daily for completed surveys! Friday, November 15, 13 Thank You

×