Re: [RESEND][RFC] BFQ I/O Scheduler

[Paolo Valente <paolo.valente@unimore.it>]

Avi Kivity ha scritto:
> Paolo Valente wrote:
>> Avi Kivity ha scritto:
>>> Jumping in at random, does "process" here mean task or mms_struct?  
>>> If the former, doesn't that mean that a 100-thread process can 
>>> starve out a single-threaded process?
>>>
>>> Perhaps we need hierarchical io scheduling, like cfs has for the cpu.
>>>
>> Hierarchical would simplify isolating groups of threads or processes.
>> However, some simple solution is already available with bfq. For 
>> example, if you have to fairly share the disk bandwidth between the 
>> above 100 threads and another important thread, you get it by just 
>> assigning weight 1 to each of these 100 threads, and weight 100 to 
>> the important one.
>
> Doesn't work.  If the 100-thread process wants to use just on thread 
> for issuing I/O, it will be starved by the single-threaded process.
>
> [my example has process A with 100 threads, and process B with 1 
> thread, not a 101-thread process with one important thread]
>
Right. I was thinking only about the case where all the 101 threads 
concurrently access the disk, and I just wanted to say that weights may 
offer more help than priorities in simple cases as this one.
Apart from this, automatically recomputing weights as needed is most 
certainly a worse solution than hierarchical scheduling.

Paolo