on the network performance of amazon s3 cloud-storage...
TRANSCRIPT
On the Network Performance ofAmazon S3 Cloud-storage Service
Valerio Persico*, Antonio Montieri◊, Antonio Pescapè*,◊
*University of Napoli “Federico II” (Italy)◊NM2 srl (Italy)
IEEE CloudNet 2016
Pisa, Italy, October 4th 2016
An increasing number of services and applications rely on Public Clouds
• XaaS: everything as a service
• Pay-as-you-go resources
• No upfront investments
• Real-time provisioning
• Ubiquitous access
• Autoscaling
Cloud infrastructures backed by huge investments from the providers
• Complex and continuously evolving infrastructures
• Hot research topic
(Public) Cloud Paradigm
COMPUTATION
STORAGE
NETWORKING
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 2
Cloud networking and traffic growth
Cloud traffic is rapidly growing
• Since 2008, most of the IP traffic has originated or terminated in a datacenter
• More than 83% of datacenter traffic will be cloud traffic by 2019**
• Public cloud is growing faster than private cloud
Cloud network performance is critical
• Without high-performance networks, there would be no such thing as cloud computing*
• The network can be the bottleneck for computation (e.g. scientific computing, video processing, etc.)
• Poor network provisioning can severely compromise user experience
*Mogul and Popa, 2012
**Cisco Global Cloud Index: Forecast and Methodology 2014-2019
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 3
Cloud storage (Storage-as-a-Service)
On-line services for archiving and backup data
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 4
Cloud network heavily impacts the Quality of Service (QoS) experienced by the customers of cloud storage services
The performance of cloud-to-user network is hard to predict
Cloud providers rarely make promises or expose details about expected performance
Simple and clean user interface available
Complexities of direct HW management abstracted away
High availability and reliability guaranteed by Service Level Agreements (SLAs)
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE
Public cloud network monitoring
Effective, efficient, and accurate monitoringis constantly needed to manage increasing complexity
• Providers rarely make promises or expose details about the network
• Customers have to cope with the limited awareness about cloud networking environments
• Performance unpredictability
• Performance variability
• Hidden datacenter topology and exact service location
5
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE
Public cloud network monitoring
Effective, efficient, and accurate monitoringis constantly needed to manage increasing complexity
• Providers rarely make promises or expose details about the network
• Customers have to cope with the limited awareness about cloud networking environments
• Performance unpredictability
• Performance variability
• Hidden datacenter topology and exact service location
6
• Non-cooperative monitoring approaches
• do not require access to any restricted information
• adopt the point of view of the general consumer
• Customers are able to• validate and extend the information
supplied by the provider
• perform informed choices among different services or different providers
Contribution
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 7
• General assessment of the performance of the service
• Evaluation of the impact on the performance of a number of configuration choices
• Usage guidelines for customers
• Large-scale distributed campaign
• Analysis of the performance of the cloud-to-user network associated with Amazon S3
Methodology
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE8
4 Cloud Regions
• North Virginia (US)• Ireland (EU)• Singapore (AP)• Sao Paulo (SA)
US
SA
EU
AP
Methodology
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE9
4 cloud regions
North Virginia (US)Ireland (EU)Singapore (AP)Sao Paulo (SA)
77 vantage points (VPs)
• United States (US)• Europe (EU)• Asia-Pacific (AP)
• Central-South America (CSA)• South Africa (ZA)
Project BISmark Platform
• Router-based• Nodes voluntarily hosted
by home users
US
CSA
EU
AP
ZA
Methodology
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE10
4 cloud regions
North Virginia (US)Ireland (EU)Singapore (AP)Sao Paulo (SA)
77 vantage points (VPs)
United States (US)Europe (EU)Asia-Pacific (AP)South Africa (SA)Central-South America (CSA)
Experimental campaign and dataset
• 7 days• 40 downloads every 2 hours• Different file sizes• Different storage classes
• Standard
• Content Delivery Network (CDN) Enabled
Methodology
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE11
4 cloud regions
North Virginia (US)Ireland (EU)Singapore (AP)Sao Paulo (SA)
77 vantage points (VPs)
United States (US)Europe (EU)Asia-Pacific (AP)South Africa (SA)Central-South America (CSA)
Experimental campaign and dataset
• 7 days• 40 downloads every 2 hours• Different file sizes• Different storage classes
• Standard
• Content Delivery Network (CDN) Enabled
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 12
Impact of the file size
• Impact on user experience
• Impact on monitoring activities• Performance depends
on the file size
• 2500KiB/s, on average when downloading 100 MiB files
• -33% w 16MiB
• -80% w 1MiB
• Goodput is impacted by TCP dynamics
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 13
Impact of the cloud region
• US region is the best global choice from both the performance and the economic point of view
• US and EU 45% better than AP and SA, on average• Counterintuitively AP and SA are associated to higher costs
US: 3562 KiB/s
EU: 2791 KiB/s
SA: 1445 KiB/s
AP: 2018 KiB/s
Impact of the placement of the VP
• Best performance obtained for US VPs retrieving objects from the same region, on average
• Worst performance observed for AP VPs retrieving objects from SA and vice versa (CSA from AP)
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 14
VPs
Impact of the placement of the VP
• Best performance obtained for US VPs retrieving objects from the same region, on average
• Worst performance observed for AP VPs retrieving objects from SA and vice versa (CSA from AP)
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 15
VPs
Impact of the placement of the VP
• Best performance obtained for US VPs retrieving objects from the same region, on average
• Worst performance observed for AP VPs retrieving objects from SA and vice versa (CSA from AP)
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 16
VPs
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 17
How much does the performance improve if users rely on buckets placed in the same geographical region?
• The performance grows by 44%, on average, considering VP and cloud bucket placed in the same geographical regions
• AP is the region with the major improvement (+77%)
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 18
Adoption of the CDN service• Better performance in terms of
goodput
• Markedly higher improvement for smaller sizes
• 100MiB +104%, on average
• 1MiB +274%, on average
• Lower performance variability• -28 % in terms of Coefficient
of Variation (CoV)
NO CDN1MiB
CDN1MiB
NO CDN100MiB
CDN100MiB
CDN
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 19
User to CDN node mapping• 38/54
CDN nodes encountered
• CDN nodes placed
• in the same geographical region of the VP
• in the US region
For only 18/48 VPs the best performing node has the highest occurrence frequency
VPs
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 20
User to CDN node mapping• 38/54
CDN nodes encountered
• CDN nodes placed
• in the same geographical region of the VP
• in the US region
For only 18/48 VPs the best performing node has the highest occurrence frequency
VPs
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 21
User to CDN node mapping• 38/54
CDN nodes encountered
• CDN nodes placed
• in the same geographical region of the VP
• in the US region
For only 18/48 VPs the best performing node has the highest occurrence frequency
VPs
CDN doesn’t lead always to better performance than S3
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 22
• +2215 KiB/s, on average (+144.11%)
• In 30% of the cases average improvement lower than 100 KiB/s
• In 119 out of 924 cases the CDN leads to worse performance
• Worst case: -43.59% degradation
Goodput gain with CDN
x
Goodput gain with CDN
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 23
ConclusionCloud-storage high level management interface hides
implementation details and performance figures
• Distributed experimental campaign leveraging the BISmark platform
• General assessment of the performance in terms of goodput
• US and EU cloud regions offer better performance (+45%, on average) at lower cost
• Enabling the CDN service leads to +144% improvement
• A number of cases found exposing a detrimental effect of the CDF service (up to -44%)
• Non-cooperative approaches- adoption of the point of view of the general customer- no need for access to information restricted to the provider
• Comprehensiveness- different providers - different use scenarios
• Predictability of experimental costs- computation cost- network-usage cost
• Results sharing- community results repository
• Ease of use- on demand measurements- no specific monitoring skills needed
• Public availabilityhttp://traffic.comics.unina.it/cloudsurf
CloudSurf a platform to monitor public-cloud networks
Up-to-date Pricing
Information
Results Repository
CloudSurfUser
Master
Cloud probe
Cloud credentials
Virtual machine
Management API
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 24
http://wpage.unina.it/valerio.persico
http://traffic.comics.unina.it/cloudsurf/
?
• For some VPs the performance is not significantly affected by the cloud region (e.g., AP12, EU7, US13)
• In other cases, a non-negligible discrepancy is measured on changing buckets (e.g., AP11, EU8)
• The globally optimal choice could be dramatically outperformed by local optimum choices
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 28
Impact of the placement of the VP
…
…
• For some VPs the performance is not significantly affected by the cloud region (e.g., AP12, EU7, US13)
• In other cases, a non-negligible discrepancy is measured on changing buckets (e.g., AP11, EU8)
• The globally optimal choice could be dramatically outperformed by local optimum choices
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 29
Impact of the placement of the VP
…
…
• For some VPs the performance is not significantly affected by the cloud region (e.g., AP12, EU7, US13)
• In other cases, a non-negligible discrepancy is measured on changing buckets (e.g., AP11, EU8)
• The globally optimal choice could be dramatically outperformed by local optimum choices
ON THE NETWORK PERFORMANCE OF AMAZON S3 CLOUD-STORAGE SERVICE 30
Impact of the placement of the VP
…
…