Accelerating Relaxation Membase Server and Fusion-io
Jun 24, 2015
Accelerating RelaxationMembase Server and Fusion-io
With a Fusion-Powered data center companies
can increase productivity, shrink time to
market, improve their customers’ experience,
and they can do it all with less hardware, less
software and less power. Fusion-io helps you
do more with less.
the industry leader
in, system, application & database
acceleration, is unleashing the potential of
performance-starved applications, allowing
companies to rethink the way they architect
their data systems.
37% of servers are massively underutilized¹…
THE DATA SUPPLY PROBLEM
LEADS TO IDLE CAPACITY
Processing performance doubles every 18 months, according to Moore's Law
4
1 Source: IDC's Server Workloads 2010, July 20102 Source: Taming the Power Hungry Data Center, Fusion-io White Paper
Server is idle
80% of the time
CPUs
Memory
Storage
Rela
tive P
erf
orm
ance
…because the performance gap continues to grow
2000 20051985 1990 1995 2010
Growing
performance
gap²
THE DATA SUPPLY PROBLEM
CAUSES DATACENTER BLOAT
Memory Rich Servers
Commodity
StorageHigh Performance
Storage
10,000 Microseconds
Storage, Networking, Servers
$ $$ $$$ $$$$ $$$$$
Software, Services
Power, Cooling, Space
Access Delay
High Performance
Networking
5
Applications StorageServers Network
Applicati
on W
ork
load
Source: Based on 3rd party forecasts and company estimates
An estimated $1.2 trillion a year is spent in the datacenter
$260 billion of that is estimated to be value wasted waiting for data
The performance of a SAN…
THE SOLUTION STARTS WITH
MINIATURIZING THE SAN
10,000 microsecond access time
…in the palm of your hand
< 30 microsecond access time
6
Source: The Storage Performance Council and Fusion-io Estimates
7
Coming Soon to a Cloud Near You
• I/O in the cloud is unreliable
• No real SLA
• AWS: ―I/O Performance: Moderate‖ – ―what?‖
• ―resources like … the disk subsystem are shared
among instances.‖
• Instances available now in Atlanta and London:
vps.net
• vps.net/wiki/using-fusion-io-on-the-cloud
Memory Rich Servers
High Performance
Storage
Storage, Networking, Servers
$ $$ $$$ $$$$ $$$$$
Software, Services
Power, Cooling, Space
Access Delay
DECENTRALIZATION
PUTS THE DATA CLOSER
8
Applications StorageServers Network
Applicati
on W
ork
load
10,000 Microseconds
High Performance
Networking
Commodity
Storage
Storage, Networking, Servers
$ $$ $$$ $$$$ $$$$$
Software, Services
Power, Cooling, Space
Access Delay
Commodity
Networking
DECENTRALIZATION
ELIMINATES WAIT AND BLOAT
9
Applications StorageServers Network
Applicati
on W
ork
load
<30 Microseconds
Commodity
Storage
Commodity Servers
ioSphere
Virtual Storage Layer (VSL)
• New extensible OS
subsystem
• Virtualizes ioMemory
ioSphere
• Purpose-built systems
management software
• Centralizes distributed
ioMemory resources
Future Applications
• Applications that
leverage VSL and
ioMemory
Fusion Monitoring
Service
• Cloud-based monitoring-
as-a-service
• To be released soon
directCache
• Selects active data to be
decentralized in
ioMemory
3rd Party Applications
• Independent 3rd party
applications such as file
systems, databases,
etc.
Internet
THE FUSION-IO LEAN COMPUTING PLATFORM
FOR JUST-IN-TIME DATA DELIVERY
IT Administrators
10
Commodity
Storage
Commodity Servers
Commodity
Networking
The value is in the software that runs
the decentralized data supply chain
Membase Server Challenge
• Persisting large key/value stores requires a
significant amount of I/O
• Membase Server quickly gets I/O bound for data
bigger than memory.
• Membase Server SQLite backend poses a challenge
for high bandwidth I/O. No direct-IO
• More cluster nodes are traditionally used to do I/O
in parallel and scale
11
Hypothesis
• For instance, by using multiple buckets, an ioDrive DUO is able to achieve >1GB/s from SQLite read/write
• Node restart time is much improved as hot data is read back into memory
• Fewer but faster cluster nodes reduce management and power requirements significantly. And much less network infrastructure is needed
12
High performance server side flash has the bandwidth and
latency to increase performance and reduce the number of
cluster nodes significantly
System Configuration
• Memcachetest was used for performance and
configuration evaluation
• System had 128GB RAM and data store was 360GB
with a fixed 20k item size. Runs done with varying
memory quotas available to Membase Server
• For comparison, 4x500GB hard drives were used in
a RAID0 configuration
• All runs were done after on-disk page bring-in
stopped by Membase Server during startup
13
Results
14
• Memcachetest setup to read data from 7 buckets, each with 1-5
million items
• Memory is pre-warmed
• All I/O is random, only limited request merging possible
• Thread count is 2-8 per bucket
Results
15
• For this test Membase Server memory quota was 100GB
• 3x the memory gives approx. 25% improvement for the ioDrive
Results
16
• Both ioDrive runs show consistent
performance after it stabilizes
• HDD 100GB run is more spikey as
all I/O requests are hugely
expensive compared to memory
Summary of Results
17
• Sets can be expensive as SQLite contention gets worse
• More buckets will improve performance compared to spindles
further
• I/O to ioDrives for this run is approx. 850MB/s reads and 150MB/s
writes
Conclusions
18
• Adding solid state storage can indeed improve I/O and allow you to
reduce number of cluster nodes
• FIO's ioMemory does improve performance compared to HDD by 40 to
50 times!
• We estimate this would translate to a 10 to 1 reduction in cluster nodes
• Reducing cluster size will reduce hw and sw costs, network
infrastructure, management overhead, power etc.
• Cluster node restart time is reduced providing better user experience
• Ability to scale more seamlessly—node at a time vs racks
• Don’t pay more for crappy instances, pay less for more intelligent ones
• Request Fusion-io to your Cloud provider
Thank you!
19
Questions?