COMP7500 Advanced Operating Systems I/O-Aware Load Balancing Techniques Dr. Xiao Qin Auburn University Spring, 2012

2 Technology Trend Big Fishes Eating Little Fishes In reality:

3 Technology Trend Supercomputer Mini- supercomputer Mainframe Mini- computer Work- station PC Massively Parallel Processors 1988 Computer Food Chain

4 Technology Trend Computer Food Chain PCWork- station Mainframe Supercomputer Mini- supercomputer Clusters Mini- computer Now who is eating whom? Server

5 Nov MEMORY BUS/CROSSBAR CPU Symmetric Multiprocessing (SMP)Massively Parallel Processor (MPP) CPU M M M M PC network cluster MPP Cluster SMP Constellations SIMD Single processor Supercomputer Trends in Top 500

6 Growth in Microprocessor Performance

7 Six Generations of DRAMs

8 Technology  dramatic change Processor –transistor number in a chip: about 55% per year –clock rate: about 20% per year Memory –DRAM capacity: about 60% per year (4x every 3 years) –Memory speed: about 10% per year –Cost per bit: improves about 25% per year Disk –capacity: about 60% per year –Total use of data: 100% per 9 months! Network Bandwidth –10 years: 10Mb  100Mb – 5 years: 100Mb  1 Gb

9 Updated Technology Trends (Summary) CapacitySpeed (latency) Logic 4x in 4 years2x in 3 years DRAM4x in 3 years2x in 10 years Disk4x in 2 years2x in 10 years Network (bandwidth) 10x in 5 years

10 I/O-intensive Applications long running simulations remote-sensing database systems biological sequence analysis

11 Motivation I/O-intensive Applications require input and output of large amounts of data. I/O performance can be a potential bottleneck. PCI Bus 264 MB/s W: 209 MB/s R: 236 MB/s Disk Write: 32 MB/s Read : 26 MB/s Memory W: 592 MB/s R: 464 MB/sC: 316 MB/s Faster! disk

12 Current Solutions Disk I/O Systems –Caching –Prefetching –Parallel I/O Limitation –Low level –Not Portable

13 Current Solutions (Cont.) Non-I/O-aware (Condor, Mosix, DQS, LSF ) Disk-I/O-awareNetwork-I/O-aware load balancing Disk-I/O Buffer Management Scheduling/Load balancing Space-sharing (PBS,Backfilling) Time-Sharing Centralized Control (PBS) Distributed Control Support Sequential Jobs Support Parallel Jobs Support Homogeneous Clusters Support Heterogeneous Clusters Coordinated Scheduling (Gang)

14 System Architecture High Bandwidth network Load Manager t1 t2 t3 disk Load Manager t3t4 disk Load Manager t5t6t7 disk Client Services Workstation 1Workstation 2Workstation n mem I/O-intensive jobs

15 Methodology I/O Intensive Applications User Specified Access Pattern Measure I/O load Predict Response Time Estimate Overhead Make Decisions Dispatch and Migration Load Balancing Schemes Data Storage Pattern

16 Outline Motivations A Disk-I/O-Aware Load Balancing Policy with Remote Execution A Disk-I/O-Aware Load Balancing Policy with Preemptive Migration Evaluation of the two Disk-I/O-Aware Policies Load Balancing for Heterogeneous Clusters Contributions and Conclusions