Multiple Sequential Reader vs Multiple Sequential Reader
===============================================
Vanilla CFQ
-----------------------------------
[Multiple Sequential Reader]                  [Multiple Sequential Reader]
nr  Max-bandw Min-bandw Agg-bandw Max-latency nr  Agg-bandw Max-latency 
16  1015KB/s  293KB/s   13705KB/s 2573K usec  1   714KB/s   1619K usec  
16  885KB/s   263KB/s   12372KB/s 2110K usec  2   1507KB/s  1890K usec  
16  833KB/s   248KB/s   11156KB/s 1924K usec  4   2276KB/s  1924K usec  
16  679KB/s   180KB/s   9351KB/s  2368K usec  8   4136KB/s  2367K usec  
16  518KB/s   149KB/s   6739KB/s  3168K usec  16  6626KB/s  3169K usec  

IO scheduler controller + CFQ
-----------------------------------
[Multiple Sequential Reader]                  [Multiple Sequential Reader]
nr  Max-bandw Min-bandw Agg-bandw Max-latency nr  Agg-bandw Max-latency 
16  513KB/s   167KB/s   7003KB/s  3415K usec  1   5826KB/s  427K usec   
16  505KB/s   169KB/s   6892KB/s  3245K usec  2   6553KB/s  334K usec   
16  514KB/s   123KB/s   6939KB/s  3457K usec  4   6556KB/s  1013K usec  
16  495KB/s   130KB/s   6808KB/s  3148K usec  8   6591KB/s  1648K usec  
16  502KB/s   124KB/s   6717KB/s  3132K usec  16  6481KB/s  3132K usec  

dm-ioaband + CFQ
-----------------------------------
[Multiple Sequential Reader]                  [Multiple Sequential Reader]
nr  Max-bandw Min-bandw Agg-bandw Max-latency nr  Agg-bandw Max-latency 
16  944KB/s   278KB/s   11020KB/s 1468K usec  1   1749KB/s  1438K usec  
16  809KB/s   305KB/s   9110KB/s  1456K usec  2   3709KB/s  1330K usec  
16  834KB/s   228KB/s   8162KB/s  1751K usec  4   4941KB/s  1717K usec  
16  730KB/s   179KB/s   8367KB/s  2098K usec  8   5485KB/s  2095K usec  
16  630KB/s   166KB/s   7197KB/s  2851K usec  16  6446KB/s  2795K usec  

Note:
- For the samples of 1, 2 and 4 sequential readers, the max latencies should
  have been low but these are still high. This is probably due to the fact
  that CFQ does not know about groups and it will still run random reader
  queue in round robin manner.

  If we do excessive throttling to improve latencies, then throughput will
  suffer because every sequential queue will get small slice and move on
  to next queue hence increasing number of seeks.