SyNS'13 Session 1: Cloud

Next Stop, the Cloud: Understanding Modern Web Service Deployment in EC2 and Azure
Keqiang He, Aditya Akella

A global, in-depth understanding of lass cloud use. Questions:
1. Who is using public clouds? (cloud-using domains, traffic profile)
2. How are Web services using the cloud (impact of failure etc)

Datasets:  University packet capture and Alexa subdomain DNS records

Domain: 4% domains are using cloud, among those, 90% are using Amazon cloud.
Traffic volume: 81.7% go to Amazon, and 18.3% to WIndows Azure
Frond End: 8% PaaS front end, 4% Load Balancer frond end, 72% VM front end
Regions: 97% of the domains are using just one region ---> single region failure can take down a large fraction of cloud-using services
Availability Zone: 33% domains are using 1 availability zone, 40% are using two.

Highlights:
1. 4% popular services using EC2/Azure
2. VMs are most popular EC2 front end, value-add features used b top domains
3. LImited zone and region usage.

Question:
Q:Dropbox is dominating so much. If Dropbox is taken out of the picture, would the result still be the same?
A: For flow analysis, I don't know. For domain analysis, it would be the same.



Viewbox: Integrating Local File Systems with Cloud Storage Services
Yupu Zhang, Remzi Aparci-Desseau

Is your data really safe?  Not really in Dropbox even if you have many copies...
Corruption populate to other copies.
Crash inconsistency ends up everywhere
"Out-of-sync" synchronization: doesn't handle client crashes correctly.
Causality incosistency: files uploaded out of sync, and cloud state won't match an FS epoch.

Why:
Semantic gap between file systems and cloud. These are designed separtely.

Viewbox: view-based synchronization
Views: in-memory snapshots of the synchronizatin folder, crated at FS epoches (e.g, journal commit time)
One active view (to get FS changes), one frozen view (to sycn client),

Compontes of Veiwbox:
1. ext4-checksum  ---> corruption detection and inconsistency detection (like transation didn't commit)
2. ViewFS (an in-memor extention to ext4-checksum. Keeps an operation log to all the metadata, so it could later appy it to frozen view?)
3. Cloud-aid recovery.

Stratos: A Network-Aware Orchestration Layer for Virtual Middleboxes in the Cloud
Anand, Aditya

(This is the middlebox scaling Aditya was talking about last year)
State of the Art: in order to steer traffic to middles boxes, tenants need to overlay network on top of cloud network

Stratos:
Forwarding controller: tranlate policies of middlebox chains to forwarding rules.
    1. Split chains into subchains at mangling/connection-terminating middle boexies
    2. Calculate forwarding rules.
Data plane
Resource controller:
     Monitor end-to-end performance of chains, then decide resources on each chain. (Use flow distribution, horizaontal scaling, instance migration and scaling up middlebox, from lighter-weighted to heavier-weighted)