Cost-Benefit Analysis of
Cloud Computing versus
Desktop Grids
Derrick Kondo, Bahman Javadi, Paul Malécot, Franck Cappello INRIA, France
David P. Anderson UC Berkeley, USA
Cloud Background
•
Vision•
Hide complexity of hardware andsoftware management from a user by offering computing as a service
•
Benefits•
Pay as you go•
Scale up or down dynamically•
No hardware management, less software managementPlatform Performance vs. Costs
Cost Performance low low high highPlatform Performance vs. Costs
Cost Performance Cluster Grid Supercomputer low low high highPlatform Performance vs. Costs
Cost Performance Cluster Grid Supercomputer low low high high Desktop Grids (DG)Platform Performance vs. Costs
Cost Performance Cluster Grid Supercomputer low low high high Desktop Grids (DG)Platform Performance vs. Costs
Cost Performance Cluster Grid Supercomputer low low high high Desktop Grids (DG)Cloud What is the relationship? DG + Cloud?
Outline
•
Performance tradeoffs•
Monetary tradeoffs•
Client hostingApples to Apples
Loosely-coupled, high-throughput, compute-intensive applications Tightly-coupled, data-intensive real-time applicationsApples to Apples
Loosely-coupled, high-throughput, compute-intensive applications Tightly-coupled, data-intensive real-time applicationslow complexity high complexity
Apples to Apples
Loosely-coupled, high-throughput, compute-intensive applications Tightly-coupled, data-intensive real-time applicationslow complexity high complexity
DG’s
Apples to Apples
Loosely-coupled, high-throughput, compute-intensive applications Tightly-coupled, data-intensive real-time applicationslow complexity high complexity
DG’s
Clouds
Comparison assuming
Method
•
Use real performance measurements•
Exported BOINC (middleware for desktop grids) project data•
Use real costs•
Large/small BOINC projects (SETI@home / XtremLab)Stages of Project & Application
Platform Construction
Stages of Project & Application
Platform
Stages of Project & Application
Platform
Stages of Project & Application
Platform
Platform Construction Application Deployment Application Execution Application Completion How long before I
Platform Construction Application Deployment Application Execution Application Completion How long before I
get X TeraFLOPS? 0 5 10 15 20 25 30 103 104 105 106
Months for registration
Number of cloud nodes
0 5 10 15 20 25 30
101 102 103
Platform Construction Application Deployment Application Execution Application Completion How long before I
get X TeraFLOPS? 0 5 10 15 20 25 30 103 104 105 106
Months for registration
Number of cloud nodes
0 5 10 15 20 25 30
101 102 103
TeraFLOPS
Can get over 20 TeraFLOPS within 6 months
Platform Construction Application Deployment Application Execution Application Completion How long before I
get X TeraFLOPS? 0 5 10 15 20 25 30 103 104 105 106
Months for registration
Number of cloud nodes
0 5 10 15 20 25 30
101 102 103
TeraFLOPS
Can get over 20 TeraFLOPS within 6 months
Strategy:
Add to BOINC project list Press releases
Forum Announcements Google Ad Sense
Respond to users (leverage volunteers)
Platform Construction Application Deployment Application Execution Application Completion
How long to deploy my batch of tasks
needing
Platform Construction Application Deployment Application Execution Application Completion
How long to deploy my batch of tasks
needing
faster response time?
!"# !"$ !"% !"!! !"" !"! !"& !"' ()*+,-./0.1/2)34,,-.3/5,6 7,82/9*,34.4:*,.;*:3)4,6< . . !"".4=6>6 !""".4=6>6 !"""".4=6>6
Platform Construction Application Deployment Application Execution Application Completion
How long to deploy my batch of tasks
needing
faster response time?
!"# !"$ !"% !"!! !"" !"! !"& !"' ()*+,-./0.1/2)34,,-.3/5,6 7,82/9*,34.4:*,.;*:3)4,6< . . !"".4=6>6 !""".4=6>6 !"""".4=6>6 For 1000 tasks,
Platform Construction Application Deployment Application Execution Application Completion
How long to deploy my batch of tasks
needing
faster response time?
!"# !"$ !"% !"!! !"" !"! !"& !"' ()*+,-./0.1/2)34,,-.3/5,6 7,82/9*,34.4:*,.;*:3)4,6< . . !"".4=6>6 !""".4=6>6 !"""".4=6>6 For 1000 tasks,
~10 minutes with 105 hosts
Strategy:
Specify lower latency bounds
Platform Construction Application Deployment Application Execution Application Completion
Platform Construction Application Deployment Application Execution Application Completion How many volunteer nodes are equivalent to 1 cloud node?
Platform Construction Application Deployment Application Execution Application Completion How many volunteer nodes are equivalent to 1 cloud node?
Platform Construction Application Deployment Application Execution Application Completion How many volunteer nodes are equivalent to 1 cloud node?
2.8 active volunteer hosts per 1 cloud node.
(Total performance still
Platform Construction Application Deployment Application Execution Application Completion How many volunteer nodes are equivalent to 1 cloud node?
2.8 active volunteer hosts per 1 cloud node.
(Total performance still
orders of magnitude better)
Strategy:
Use statistical prediction of
Platform Construction Application Deployment Application Execution Application Completion
Platform Construction Application Deployment Application Execution Application Completion How long should I wait for task completion?
Platform Construction Application Deployment Application Execution Application Completion How long should I wait for task completion?
Platform Construction Application Deployment Application Execution Application Completion How long should I wait for task
completion? for 3.7 hour work unit (on 3GHz host).Median project latency bound: 9 days
Ratio of lat. bound / exec time > 5. Good success rates: 96.1%
Platform Construction Application Deployment Application Execution Application Completion How long should I wait for task completion? Strategy:
See BOINC Catalog for typical deadlines and
compute/comm/mem ratios.
Median project latency bound: 9 days for 3.7 hour work unit (on 3GHz host).
Ratio of lat. bound / exec time > 5. Good success rates: 96.1%
Monetary Tradeoffs
•
Client hosting on cloud•
Not worth it and never will•
Server hosting on the cloudWhat about Client
Cost of Clouds versus DG
How long until DG’s more cost effective than Clouds?Cost of Clouds versus DG
! " #! #" $! $" %! #!!#& #!!#" #!!#' #!!#% #!!#$ #!!## ()*+,-./01,/230456 7,8./9,)/:;<( / / =.),>;0? @ABCDE*>, A7$ F#'!'/G*HI8.,,)/8*3,5 2'J&/-H*I3/8*3,56 '"&$/G*HI8.,,)/8*3,5 2#&#$/-H*I3/8*3,56 How long until DG’s more cost effective than Clouds?Cost of Clouds versus DG
! " #! #" $! $" %! #!!#& #!!#" #!!#' #!!#% #!!#$ #!!## ()*+,-./01,/230456 7,8./9,)/:;<( / / =.),>;0? @ABCDE*>, A7$ F#'!'/G*HI8.,,)/8*3,5 2'J&/-H*I3/8*3,56 '"&$/G*HI8.,,)/8*3,5 2#&#$/-H*I3/8*3,56 How long until DG’s more cost effective than Clouds? After ~13 daysWhat are total costs over time?
0 5 10 15 20 0 1 2 3 4 5 6 7 8x 10 4
Platform age (days)
Total cost EC2 100n EC2 1000n EC2 10000n XtremLab SETI@home What are total costs over time?
0 5 10 15 20 0 1 2 3 4 5 6 7 8x 10 4
Platform age (days)
Total cost EC2 100n EC2 1000n EC2 10000n XtremLab SETI@home What are total costs over time?
Cost of Clouds versus DG (2)
After 3 days, DG is cheaper
Cost of Clouds versus DG (3)
How many months of DG can
X months of Cloud buy me?
0 5 10 15 20 25 0 500 1000 1500 2000 2500 3000 3500
Months with EC2
Months with VC
Xtremlab / 1000n EC2 SETI@home / 1000n EC2 Xtremlab / 10000n EC2 SETI@home / 10000n EC2
Cost of Clouds versus DG (3)
How many months of DG can
X months of Cloud buy me?
0 5 10 15 20 25 0 500 1000 1500 2000 2500 3000 3500
Months with EC2
Months with VC
Xtremlab / 1000n EC2 SETI@home / 1000n EC2 Xtremlab / 10000n EC2 SETI@home / 10000n EC2
Cost of Clouds versus DG (3)
How many months of DG can
X months of Cloud buy me?
12 months of Cloud
can pay for > 125 DG years. DG way
Equivalent Clouds given BOINC Costs
Given BOINC costs, what size Cloud can I buy?
Equivalent Clouds given BOINC Costs
Given BOINC costs, what size Cloud can I buy?
Equivalent Clouds given BOINC Costs
Given BOINC costs, what size Cloud can I buy?
1.75 TeraFLOPS, 7 TeraBytes
.75 TeraFLOPS, 3 TeraBytes
Equivalent Clouds given BOINC Costs
Given BOINC costs, what size Cloud can I buy?
2 orders of magnitude
lower than BOINC
1.75 TeraFLOPS, 7 TeraBytes
.75 TeraFLOPS, 3 TeraBytes
What about Server
Hosting on the Cloud?
Number of Hosts over Time
Load variation exits w/ publicity,
projects run out of work, etc.
Clouds take care of server
management. Potential to
exploit clouds.
Server Costs on a Cloud
How much to host
Server Costs on a Cloud
How much to host
Server Costs on a Cloud
How much to host
BOINC server on cloud?
Server Costs on a Cloud
1.7 - 16.5 times cheaper to host on Cloud,
but bandwidth is expensive.
How much to host
BOINC server on cloud?
Cloud Resources with Given Budget
How big of a server can I support with given budget?
Cloud Resources with Given Budget
How big of a server can I support with given budget?
Cloud Resources with Given Budget
How big of a server can I support with given budget?
Cloud Resources with Given Budget
Many project servers
are sustainable on the cloud
How big of a server can I support with given budget?
Summary
•
Performance tradeoffsSummary
•
Performance tradeoffs•
20 DG TeraFLOPS within 6 months•
Monetary tradeoffs•
Client hosting•
After 13 days, DG more cost effective•
Server hosting•
DG server on cloud is cost-effective•
Best for small-medium sized projects•
Savings of at least 40%Summary
•
Performance tradeoffs•
20 DG TeraFLOPS within 6 months•
Monetary tradeoffs•
Client hosting•
After 13 days, DG more cost effective•
Server hosting•
DG server on cloud is cost-effective•
Best for small-medium sized projects•
Savings of at least 40%• http://mescal.imag.fr/membres/derrick.kondo/cloud_calc.xlsx