locks and kernel randomness...

[Harrison Grundy]

On 02/24/15 23:58, Alfred Perlstein wrote:
> 
> On 2/24/15 7:23 PM, John-Mark Gurney wrote:
>> K. Macy wrote this message on Tue, Feb 24, 2015 at 15:33 -0800:
>>>> If someone does find a performance issue w/ my patch, I WILL
>>>> work with them on a solution, but I will not work w/ people
>>>> who make unfounded claims about the impact of this work...
>>>> 
>>> <shrug> ... The concerns may be exaggerated, but they aren't 
>>> unfounded. Not quite the same thing, but no one wants to spend
>>> the
>> Till someone shows me code in the kernel tree where this is even
>> close to a performance problem, it is unfounded...  I've asked,
>> and no one has
>> 
>>> cycles doing a SHA256 because it's "The Right Thing"(tm) when
>>> their use case only requires a fletcher2.
>> Depends upon what you're doing.. I haven't proposed changing
>> ZFS's default to sha256, so stop w/ the false equivalences...
>> 
>>> If it doesn't already exist, it might also be worth looking in
>>> to a more scalable CSPRNG implementation not requiring locking
>>> in the common case. For example, each core is seeded separately
>>> periodically so that has a private pool that is protected by a
>>> critical section. The private pool would be regularly refreshed
>>> by cpu-local callout. Thus, a lock would only be acquired if
>>> the local entropy were depleted.
>> I'm not discussing this until you read and reply to my original
>> email, since it's clear that my original email's contents has
>> been ignored in this thread...
>> 
> What is final proposal?  More spinlocks?  That is not a good idea.
> 
> Doing a single buildworld is not enough.  Ask netflix or someone
> with a real load of 1000s of threads/processing to do testing for
> you if you truly want to touch scheduler.

sched_ule runs this code once every .5 to 1.5 seconds, depending on
the value of random, so using a CSPRNG there wouldn't actually be
noticeable. (We're talking about a few thousand cycles, when the
existing implementation has to make a remote memory read/write
numpackages-1/numpackages percent of the time, which costs tens of
thousands of cycles. Switching to a per-CPU CSPRNG is actually faster
in those cases.)

That being said, I believe the plan is to remove random() from
sched_ule entirely. It doesn't need it to perform the balancing, and
we can just use the LCG from cpu_search, if get_cyclecount isn't viable.

--- Harrison