search

uit-inf-3203 / Course

Course material and information for the INF-3203 course at UiT
5 stars 0 forks source link

Questions for Windows Azure Storage #13

Closed fjukstad closed 9 years ago

fjukstad commented 9 years ago

Hey @uit-inf-3203/students

Here are Ruben's questions for his presentation next week:

  1. WAS try to keep storage stamps around 70% utilized in terms of capacity, transactions and bandwidth. This is because they want to keep at least 20% available in reserve for: A: disk short stroking to gain better seek time and higher throughput by utilizing the outer tracks of the disk B: Continue providing storage capacity and availability in the presence of a rack failure within a stamp. In the case of a rack failure, 20% of the spared storage must be utilized. How big impact, if any, will this have to that namespace in terms of performance and throughput? Peter Haro
  2. WAS issue reads within an extent with a "deadline" which specifies that the read should not be attempted if it cannot be fulfilled within that time constraint. WAS have three replications of an extent node. If none of these three extent nodes can fulfill the deadline constraint, WAS immediately issue a reconstruction of the fragment from an erasure coded extent. Is this solution suitable? How bad will this violate the time constraint, if anything at all? Håvard Hemmingsen Johansen
  3. Each account using WAS are limited to only 100 terraBytes. This implies that large customers who want to store petabytes of data, must have several accounts in order to store their data. Could WAS offer an abstraction where one account may be the "primary account" which is in control of several "sub accounts", so the customer only have to communicate with what it thinks is one single account? Håvard Emanuel Hamre Mathisen
  4. WAS use range partitions because performance isolation becomes easier when given account objects are stored together within a set of range partitions. The paper states that hashing would give better workload between servers, but lose the locality of objects for isolation and enumeration. Would it be possible to achieve locality of objects with a tailored hashing scheme? Eirik Iversen Mortensen
  5. According to the paper, the CAP theorem have been overcome in WAS. The stream layers simple append-only model secures high availability, and the partition layer (which is built upon the stream layer), provides strong consistency. This is only within a geographical region, so is the CAP theorem then overcome across geographic regions? Christian Bergvoll Vik

Bjørn