Skip to content

Latest commit

 

History

History

2017_06_30_scheduler_results

README.md

1_ I/O benchmarks for evaluation

  • explains the four cases of scenarios.
  • each benchmark contains a number of tasks issuing I/O requests of different types and sizes

1 read (running alone) with 4KB

  • filename: 1_4_read.fio
[read_one_4]
filename=/dev/sdb
bs=4k
numjobs=1
direct=1
rw=read
runtime=20

1 write (running alone) with 4KB

  • filename: 1_4_write.fio
[write_one_4]
filename=/dev/sdb
bs=4k
numjobs=1
direct=1
rw=write
runtime=20

1 read (running alone) with 128KB

  • filename: 1_128_read.fio
[read_one_128]
filename=/dev/sdb
bs=128k
numjobs=1
direct=1
rw=read
runtime=20

1 write (running alone) with 128KB

  • filename: 1_128_write.fio
[write_one_128]
filename=/dev/sdb
bs=128k
numjobs=1
direct=1
rw=write
runtime=20

1 read and 1 write with 4KB

  • filename: 1_1_4_read_write.fio
[global]
filename=/dev/sdb
runtime=20
direct=1

[reader_4]
bs=4k
numjobs=1
rw=read

[writer_4]
bs=4k
numjobs=1
rw=write

16 read and 16 write with 4KB

  • filename: 16_4_16_4_read_write.fio
[global]
filename=/dev/sdb
runtime=20
direct=1

[reader_4]
bs=4k
numjobs=16
rw=read

[writer_4]
bs=4k
numjobs=16
rw=write

16 read of 4KB and 16 read of 128KB

  • filename: 16_4_16_128_read.fio
[global]
filename=/dev/sdb
runtime=20
direct=1

[reader_4]
bs=4k
numjobs=16
rw=read

[reader_128]
bs=128k
numjobs=16
rw=read

16 write of 4KB and 16 write of 128KB

  • filename: 16_4_16_128_write.fio
[global]
filename=/dev/sdb
runtime=20
direct=1

[writer_4]
bs=4k
numjobs=16
rw=write

[writer_128]
bs=128k
numjobs=16
rw=write

2_ Evaluation of Task fairness

  • use proportinoal slowdown
  • When n tasks compete for I/O simultaneously, equal resource sharing suggest that each task should experience a factor of n slowdown compared to running alone.
  • When all tasks experience better performance than the proportional slowdown, we further measure fairness according to the slowdown of the slowest tasks.
  • Specifically, scheduler S1 achieves better fairness than scheduler S2 if the slowest task under S1 makes more progress than the slowest task under S2.

- implementation

- ( avg of I/O of concurrent tasks / I/O of one task running alone )
- ( I/O performance of the slowest task / I/O of one task running alone )

3_ Evaluation of Responsiveness

- implementation

- ( avg of worst case (99.9 percentile) I/O request )