Speed Up Your LAMP Stack Applications With Dolphin Express Presentation

Improve Performance and save Money,
Space and Power with Dolphin for

MySQL based Web Servers
Copyright 2008. All rights reserved. 1
Topics
!Dolphin Quick Facts

!Dolphin Express for Database Clusters
– Superior TCO
– Performance
– Fault Tolerance
!Q&A

Quick Facts about Dolphin
!Global provider of interconnect solutions since 1994

– Main locations Oslo, Norway and Massachusetts, USA
!Leading Products and Technology

– Dolphin Express
! Enterprise Clustering and I/O Expansion
! Embedded OEM

Dolphin Express for
Database Clusters
TCO, Performance & Fault
Tolerance

Introducing Dolphin Express
Dolphin Express provides the industry’s best solution for

maximizing the performance of every database
application.
! 4X+ for Oracle
! 4X+ for MySQL
! Dramatically increase concurrent user capacity
! Substantially lower cost than alternatives
! No changes to applications or databases
! Transparent and automatic High Availability

Dolphin Express™
! Integrated Solution
– Dolphin Express Hardware
– SuperSockets™ Software
! Transparent to Applications
– NO changes to applications necessary
– ALL Applications using Sockets API (TCP-UDP-RDS) will work
– Use of LD_PRELOAD
! Scalable
– Scales to hundreds of nodes with 2-D or 3-D distributed switching
! High Availability - Fault Tolerance
– NO single point of Failure
– Automatic rerouting in case of node failures
– Multi-Channel Support with Automatic Channel Bonding
– Automatic Fail-Over to Ethernet
! Low TCO with Investment Protection
– Software compatibility with future hardware products

Lowest Cost and Best Performance
With Dolphin Without Dolphin

Qty Desc Price Qty Desc Price
6 Servers $ 45,000 12 Servers $ 90,000
6 RAC licenses $120,000 12 RAC licenses $240,000
6 Dolphin cards $ 6,000 12 Ethernet cards $ 1,800
1 Support (15/20) $ 30,750 1 Support (15/20) $ 61,500
Total: $201,750 Total: $393,300
• Dolphin solution is $191,550 (49%) less than competitor’s

• Dolphin solution is $65,706 less per year in OpEx (source: innovate.intel.com)
• 3 Year TCO savings: $354,000
Conclusion: Dolphin delivers the industry’s fastest solution for applications, while dramatically
reducing the Capital and ongoing Operational Expenses associated with those environments.

Dolphin Express for
Database Clusters
Tolerance

Web Server Requirements
!Adequate Capacity
– Scale with Traffic
! Throughput
! Response Time
!Reliable
– No Downtime visible to Clients
!Cost Efficient

DBT2 Message Size Distribution
Message Size Histogram

%
80
70
60
50
40
30
20
10
0
<64 <256 <1024 <4096 <16384 >=16384
Message size (Bytes)
Data from DBT2 Benchmark
by Mikael Ronström

Oracle RAC Message Sizes
Oracle inter-node Communication
600000
Updater 16k
Reader 16k
500000 Updater 8k
Reader 8k
NO transfers
Number of Messages
400000
larger than
300000 8k/ 16k blocks
(16k = max.
200000
cache block size)
100000
0
0<64 64<256 256<1k 1k<4k 4k<16k 16k<64k 64k<
Transfer Size

Latency
Transactions consist of Short Messages

Client
Servers
Request
Response
Transaction and other

Database applications:
Latency is Key for Response Time
and Transaction Throughput

Latency and Bandwidth
Low Latency…
- High Bandwidth

Latency Fundamentals
Packet sent Packet received

from application by application
Data Packet Data Packet
Transmission time tp Latency tl

16-bytes Packet Latency

Interconnect for DB Clusters
!Database communication patterns consist of short

data packets
– Performance is sensitive to latency and overhead
! Response time
! Throughput
!Batching of messages
– Improves throughput to a certain degree, but
– Creates longer response times
– Introduces more overhead

GigE/10GigE Network Stack
User level Application

!1 CPU Cycle
Operating System Socket Library per bit
TCP/IP Requires 1 proc. @10GHz (!)
for full bandwidth
Device Driver
Network Interface
10GigE Switch 20 - 50µs with 10GigE
Network Interface
Device Driver Network time @256Bytes payload
TCP/IP GigE : 3.12µs
10GigE: 0.31µs
Operating System Socket Library Diff.: 2.81µs
35µs - 2.81µs = 32.19µs (8%)

Dolphin Express Network Stack

!1 CPU Cycle
Operating System Sockets Library per 64 bits
SuperSockets
Dolphin Express
Interface
Fault Tolerant <3µ
<3µs is ! 10x faster than Ethernet
Interconnect Fabric
Interface
Dolphin Express
SuperSockets
Sockets Library Operating System

Shared-Nothing Data Transfers (Ether)
This is how other interconnects work:

Application
System
Network Adapter
6
1 2 3 4 5
User System 1 Application process stores data to memory

Data Buffer 2 3 System process copies data to system buffer
Host Memory 4 Protocol Processing
5 6 Data is copied to network adapter
n Bus data transfer (usually by a DMA engine)

Shared Address Space Data Transfers
This is how Dolphin Express does it:
Application Application
System System
1
Network Network
Adapter Adapter
User System System User

Data Buffer Buffer Data
Host Memory Host Memory
1 Application process stores data directly to remote memory

Dolphin SuperSockets™ Overview
!Berkeley Sockets API on Dolphin Express

!TCP/UDP/RDS compatible
!Data transfer through remote shared memory
!Flexibility through Configuration Files
– Specifying Cluster Nodes
– Specifying Communication Ports
!IP Address Virtualization

SuperSockets Protocol Handling
!RMA default limits

– In-line Protocol < 1kBytes
– Short Protocol 1kBytes < 8KBytes
– Long Protocol 8kBytes <
!RDMA
– Selectable
! Per Port
! Lower transfer limit selectable per application
!All tuning parameters selectable per application

– Knowledge based defaults

Removing Database Cluster Bottlenecks
!Available CPU cycles

– Network Overhead
– TCP processing eats CPU cycles
– Dolphin Express with SuperSockets Removes TCP/IP
overhead - Reducing CPU Load
! Dolphin Hardware provides built-in data integrity
!Network Latency
– Database Traffic consists of short data packets
!Network Efficiency
– SuperSockets utilizes Remote Memory Access (RMA)
! Reduces CPU overhead to transfer data to remote node
! Important for short transfers
! Reduces Latency

Removing Database Cluster Bottlenecks (2)
!Interrupt Processing
– Dolphin Express has integrated Checksums
! No Interrupt processing for Network Protocol
– Interrupt only required to wake sleeping process waiting
for events on the Dolphin Express socket
– Per Socket Adaptive cross-over between polling and
interrupt
!High Bandwidth with low overhead

– Dolphin Express provides
! Remote Direct Memory Access (RDMA)
! Reduces CPU overhead for large transfers

Oracle RAC Response Time
Oracle Block Service Response Time

16k Block Size
5
4,5
4
3,5
Milliseconds
3
2,5
2
1,5
1
0,5
0
1 2 3 4 6 8 10 12 14 16 20 24 28 32
Number of Clients
Dolphin Express Gigabit Ethernet

Oracle RAC Block Transfer Rate
Oracle Block Service Rate

16k Block Size
30000
25000
Blocks per Second
20000
15000
10000
5000
0
1 2 3 4 6 8 10 12 14 16 20 24 28 32
Number of Clients
Dolphin Express Gigabit Ethernet

TPC-W Benchmark Set-up
Simulated Webserver
Application Server (Apache Tomcat)
MySQL server(s)
NDB Data nodes

TPC-W Benchmark Cluster Set-up
Application Web Server

Server
MySQL MySQL MySQL MySQL

server server server server
NDB NDB NDB NDB

Data Data Data Data
Node Node Node Node
Machine 1 Machine 2

MySQL LAMP Stack TPC-W Response Time
Mean Response Time - 6 Nodes/6MySQL Servers
80,00
70,00
60,00
Milliseconds
50,00
40,00
30,00
20,00
10,00
-
4 8 16 32 64 96
Threads
Dolphin Express GigaBit Ethernet

MySQL LAMP Stack TPC-W Throughput
Throughput - 6 Nodes/6 MySQL Servers
70
60
Webinteractions/s
50
40
30
20
10
0
4 8 16 32 64 96
Threads
Dolphin Express GigaBit Ethernet

Dolphin Express for
Database Clusters
Tolerance

Two Interconnect Fabrics - Normal Operation
Dolphin Express
Failover scenarios:
! A node in the cluster fails
! A network card fails
Ethernet

2-D Dataflow
CPUs
Caches
Bridge Dolphin
2-D Card
Memory
Memory CPUs
Memory
Memory
Caches
CPUs Bridge Dolphin

2-D Card
Caches
Memory
Memory
Memory
Memory
Bridge Dolphin
2-D Card
Memory
Memory
Memory
Memory
Request
Response

Failing Node
Dolphin Express
Ethernet

2-D Dataflow w/failing Node
Memory
Memory
Memory
Memory
CPUs
Dolphin
Caches 2-D Memory
Memory
Memory
Memory
Memory CPUs
Memory
Memory Dolphin
Memory Caches 2-D
CPUs
Dolphin
Caches 2-D
Request
Response

Failing Network Card
Dolphin Express
Dolphin SuperSockets
Automatically
fails-over
to Ethernet
Ethernet

Questions &
Answers

Thank you for attending!
!Contact Dolphin:
– Einar Rustad: er@dolphinics.com
– John Longmire: jl@dolphinics.com
– Hugo Kohmann: hugo@dolphinics.com
– US switchboard: 1 508 786 9950

– www.dolphinics.com

Speed Up Your LAMP Stack Applications With Dolphin Express Presentation

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Speed Up Your LAMP Stack Applications With Dolphin Express Presentation

Uploaded by

Copyright:

Available Formats

Improve Performance and save Money,

Space and Power with Dolphin for

!Dolphin Quick Facts

Copyright 2008. All rights reserved. 2

!Global provider of interconnect solutions since 1994

!Leading Products and Technology

Copyright 2008. All rights reserved. 3

Copyright 2008. All rights reserved. 4

Dolphin Express provides the industry’s best solution for

Copyright 2008. All rights reserved. 5

Copyright 2008. All rights reserved. 6

With Dolphin Without Dolphin

• Dolphin solution is $191,550 (49%) less than competitor’s

Copyright 2008. All rights reserved. 7

Copyright 2008. All rights reserved. 8

Copyright 2008. All rights reserved. 9

Message Size Histogram

Copyright 2008. All rights reserved. 10

Oracle inter-node Communication

Copyright 2008. All rights reserved. 11

Transactions consist of Short Messages

Transaction and other

Copyright 2008. All rights reserved. 12

Copyright 2008. All rights reserved. 13

Packet sent Packet received

Data Packet Data Packet

Transmission time tp Latency tl

Copyright 2008. All rights reserved. 14

Copyright 2008. All rights reserved. 15

!Database communication patterns consist of short

Copyright 2008. All rights reserved. 16

User level Application

Copyright 2008. All rights reserved. 17

User level Application

User level Application

Copyright 2008. All rights reserved. 18

This is how other interconnects work:

User System 1 Application process stores data to memory

Copyright 2008. All rights reserved. 19

This is how Dolphin Express does it:

User System System User

1 Application process stores data directly to remote memory

Copyright 2008. All rights reserved. 20

!Berkeley Sockets API on Dolphin Express

!IP Address Virtualization

Copyright 2008. All rights reserved. 21

!RMA default limits

!All tuning parameters selectable per application

Copyright 2008. All rights reserved. 22

!Available CPU cycles

Copyright 2008. All rights reserved. 23

!High Bandwidth with low overhead

Copyright 2008. All rights reserved. 24

Oracle Block Service Response Time

Copyright 2008. All rights reserved. 25

Oracle Block Service Rate

Copyright 2008. All rights reserved. 26

Application Server (Apache Tomcat)

NDB Data nodes

Copyright 2008. All rights reserved. 27

Application Web Server

MySQL MySQL MySQL MySQL