StarFish : highly-available block storage

StarFish: highly-available block storage

資訊三 李益昌 B00902051資訊三 何柏勳 B00902097

Introduction Data protection

Disk failure V.S. catastrophic site failure Low price of disk drives and high-speed networking infrastructure StarFish

Survive catastrophic site failure Use IP network : (1) geographically-dispersed (2) inexpensive Good performance Block level

Architecture One Host Element(HE)

Provides storage virtualization and read cache N Storage Element(SE)

Q: write quorum size. Synchronous updates to a quorum of Q SEs, and asynchronous updates to the

rest. Communicate by TCP/IP over high speed network

Architecture Recommended configuration

N=3 Q=2

Architecture Another configuration

Data consistency and SE recovery Log

sequential number NVRAM

Data consistency Failure

RAID or network connection fails SE recovery

Quick recovery Replay recovery Full recovery

Availability and reliability analysis Parameter

SE failure process : SE recovery process : Number of SEs : N Quorum size : Q

Model SEs failure process is i.i.d Poisson process with mean rate SEs recovery process is i.i.d Poisson process with mean rate HE failure process Poisson process with mean rate HE recovery process Poisson process with mean rate

Availability a HE or SE is available if it can serve data Availability of StarFish A(Q, N) : the steady-state probability that at least Q

SEs are available

is called load , Repairman model

Availability(cont.)

Availability(cont.)

SE availability = 1- X ★ 9 : the number of 9s in an availability measure Fixed N, availability decreases with large Q

Trade off availability for reliability

Reliability Probability of data loss

HE and Q SEs fails The reliability increases with larger Q Two approach

Q > floor(N/2) and at least Q SEs are available Reduce availability

Read-only consistency

Read-only consistency Available in read-only mode during failure.

Read-only mode obviates the need for Q SEs to be available to handle updates. Increase availability

Availability with Read-only Consistency

Implementation

Performance measurements

Setting Gigabit Ethernet(GbE) with dummynet controlling delays and bandwith limit

to model Internet links Different network delays

1, 2, 4, 8, 23, 36, 65 ms Different bandwidth limitations

31, 51, 62, 93, 124 Mb/s Benchmark

Micro-benchmark PostMark

Effects of network delays and HE cache size

Larger cache improves performance Larger cache doesn’t change the response time of write requests

Normal Operation and placement of the far SE

Normal Operation and placement of the far SE

Normal Operation and placement of the far SE

Observation Performance is affected by two parameters

Write quorum size Q Delay to the SE

StarFish performs adequately when one of the SEs is placed in a remote location At least 85% of the performance of a direct-attached RAID

Recovery

Performance degrades more during full recovery

Conclusion The StarFish system reveals significant benefits from a third copy of data at

an intermediate distance A StarFish system with 3 replicas, a write quorum size of 2, and read-only

consistency yields better than 99.9999% availability assuming individual Storage Element availability of 99%