Racs-EV-Java. Gary Zibrat (gdz4)

Size: px

Start display at page:

Download "Racs-EV-Java. Gary Zibrat (gdz4)"

Sylvia Maxwell
5 years ago
Views:

1 Racs-EV-Java Gary Zibrat (gdz4)

2 Motivation Large organizations want to store data in the cloud (e.g. Library of Congress, Netflix, Reddit) Not only does do users pay per byte of data currently in the cloud, but also per byte of data transferred to and from.

3 Current Prices Storage Transfer out Put Request Get Request Microsoft $0.024 per GB/month $0.080 per GB $ per 1,000 transactions $ per 1,000 transactions Amazon $ per GB/month $.080 per GB $0.005 per 1,000 requests $ per 1,000 requests Google $0.026 per GB/month (flat rate) $0.080 per GB $0.01 per 1,000 requests $0.001 per 1,000 requests

4 Current Prices Example (500TB) Storage Transfer out (all data) Put Request Get Request Microsoft Amazon Google $ per year $ per year $ per year $ ~0 ~0 $ ~0 ~0 $ ~0 ~0 Moving from Amazon to Microsoft would cost roughly 2.75 years worth of storage! Large customers can t leave due to slight price increases.

5 riginal Racs EBK 1 R A EBK 2 C S

6 riginal Racs R A C E B K 1 E B K 1 E B K 2 E B K 2 S E B K 1 E B K 2

7 riginal Racs Since files are split up, cloud computation requires reassembling the files. nly part of the file may be in the same provider as where a user wishes to do cloud computation. R A C S E B K 1 E B K 1 E B K 1 E B K 2 E B K 2 E B K 2 Cloud Compute on EBK 1 Cloud Compute on EBK 2

8 Racs-EV EBK 1 R A EBK 2 EBK 3 C S

9 Racs-EV R EBK 1 A C EBK 2 S EBK 3

10 Racs-EV Now the files are still distributed evenly, but don t need to be reassembled for cloud computation which saves on transfer fees and transfer time. R A C S EBK 1 EBK 2 EBK 3 Cloud Compute on EBK 1 Cloud Compute on EBK 2 Cloud Compute on EBK 3

11 Racs-EV R EBK 1 A C EBK 2 S Parity of 1 and 2

12 Current Prices Example (625TB) Storage Transfer out (all data) Storage(625)- Storage(500) Microsoft Amazon Google $ per year $ per year $ per year $ $36864 per year $ $44544 per year $ $39936 per year Costs are for with 5 providers and the parity file turned on. Transfer out doesn t include extra 125 TB since parities aren t transfer.

13 verview - Multiple Proxies Clients C1 Repositories R1 C2 ZooKeeper R2 C3 R3 ZooKeeper is a distributed coordination system. For, it is used to get locks and reliably store meta-data.

14 -EV API PUT (Bucket, Key, Data) GET (Bucket, Key) PUTAT (Bucket, Key, Data, Repo) LCATE(Bucket, Key) DELETE (Bucket, Key) PUTS(Buckets...,Keys,Datas ) GETS(Buckets,Keys )

15 Simplified Put Clients C1 Repositories R1 C2 ZooKeeper R2 C3 R3

16 Simplified Put Clients F4 F3 P2 Repositories C1 R1 C2 ZooKeeper R2 F1 F2 P1 C3 R3 Files are grouped by size (in object groups) to reduce overhead of parity object. Group size is the same as the number of repositories.

17 Simplified Put Clients Repositories C1 R1 F4 F2 C2 ZooKeeper R2 F3 P1 C3 R3 P2 F1 Each repository will contain only one object from each group to allow for fault tolerance.

18 Problems with (-EV) String Xor for 16 MB string

19 Problems with (-EV) String Xor for 16 MB string (x) denotes number of threads

20 String Xor for 16 MB and 128 MB strings on 4 core 2 threads machine. Regular Python crashed with memory problems (Numpy did too on 32 cores)

21 Challenges with -EV Eight different objects to modify on put Data on cloud bjectgroup Parity File bjectgroup freelist (keeps track of groups with space) Key to object group mapping Previous key objectgroup (remove key from group) Previous key data on cloud (remove it) Previous parity file

22 Solution Eight different objects to modify on put Data on cloud bjectgroup Parity File bjectgroup freelist (keeps track of groups with space) Key to object group mapping Previous key objectgroup (remove key from group) Previous key data on cloud (remove it) Previous parity file

23 Solution (or not) Locking alone doesn t solve the problems Could lose connection at any point The lock could be lost at any point. if(lock.isacquired()) then modifydata() isn t atomic Similar problems to updating hard drive Things have to be done in a particular order

24 Solutions with -EV 1. Turns out this order is pretty good a. Data on cloud b. bjectgroup c. Parity File d. bjectgroup Freelist (keeps track of groups with space) e. Key to object group mapping f. Previous key objectgroup (remove key from group) g. Previous key data on cloud (remove it) h. Previous parity file

25 Solutions with -EV 1. Turns out this order is pretty good Register Intent b. Data on cloud c. bjectgroup d. Parity File e. bjectgroup Freelist (keeps track of groups with space) f. Key to object group mapping (Deregister Intent, Register Intent to delete, and part f) atomically g. Previous key objectgroup (remove key from group) h. Previous key data on cloud (remove it) i. Previous parity file

26 if(lock.isacquired()) then... This isn t atomic. Requires versioning Lock Download Note version Modify Upload with version Version same: Success Version different: Failure Unlock

27 The Data 3 Repos 3 EC2 Instances 2 cores 2 threads/core 1.7 GHz 8 GB ram Clients: 9 Clients 20 Files of each size per client Clients send then wait for response

28 The Data Part 2 3 Repos 3 EC2 Instances 2 cores 2 threads/core 1.7 GHz 8 GB ram Clients: 18 Clients 20 Files of each size per client Clients send then wait for response

29 Future Plans Always room for optimization to faster Cloud computation

30 Future Plans - Cloud Computation Send same computation request to all proxies R1 C1 R2 R3 Zookeeper not shown; Connections from each R to each not shown

31 Future Plans - Cloud Computation will see which requests belong to its repo R1 C1 R2 R3 Zookeeper not shown; Connections from each R to each not shown

32 Future Plans - Cloud Computation will see which requests belong to its repo R1 C1 R2 R3 Zookeeper not shown; Connections from each R to each not shown

33 Future Plans - Cloud Computation Client will receive back the results R1 C1 R2 R3 Zookeeper not shown; Connections from each R to each not shown

34 Demo/Questions

RACS: Extended Version in Java Gary Zibrat gdz4

RACS: Extended Version in Java Gary Zibrat gdz4 Abstract Cloud storage is becoming increasingly popular and cheap. It is convenient for companies to simply store their data online so that they don t have