lkml.org 
[lkml]   [2015]   [Jun]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/4] mm: Send one IPI per CPU to TLB flush all entries after unmapping pages

* Mel Gorman <mgorman@suse.de> wrote:

> On a 4-socket machine the results were
>
> 4.1.0-rc6 4.1.0-rc6
> batchdirty-v6 batchunmap-v6
> Ops lru-file-mmap-read-elapsed 121.27 ( 0.00%) 118.79 ( 2.05%)
>
> 4.1.0-rc6 4.1.0-rc6
> batchdirty-v6 batchunmap-v6
> User 620.84 608.48
> System 4245.35 4152.89
> Elapsed 122.65 120.15
>
> In this case the workload completed faster and there was less CPU overhead
> but as it's a NUMA machine there are a lot of factors at play. It's easier
> to quantify on a single socket machine;
>
> 4.1.0-rc6 4.1.0-rc6
> batchdirty-v6 batchunmap-v6
> Ops lru-file-mmap-read-elapsed 20.35 ( 0.00%) 21.52 ( -5.75%)
>
> 4.1.0-rc6 4.1.0-rc6
> batchdirty-v6r5batchunmap-v6r5
> User 58.02 60.70
> System 77.57 81.92
> Elapsed 22.14 23.16
>
> That shows the workload takes 5.75% longer to complete with a similar
> increase in the system CPU usage.

Btw., do you have any stddev noise numbers?

The batching speedup is brutal enough to not need any noise estimations, it's a
clear winner.

But this PFN tracking patch is more difficult to judge as the numbers are pretty
close to each other.

> It is expected that there is overhead to tracking the PFNs and flushing
> individual pages. This can be quantified but we cannot quantify the indirect
> savings due to active unrelated TLB entries being preserved. Whether this
> matters depends on whether the workload was using those entries and if they
> would be used before a context switch but targeting the TLB flushes is the
> conservative and safer choice.

So this is how I picture a realistic TLB flushing 'worst case': a workload that
uses about 80% of the TLB cache in a 'fast' function and trashes memory in a
'slow' function, and does alternate calls to the two functions from the same task.

Typical dTLB sizes on x86 are a couple of hundred entries (you can see the precise
count in x86info -c), up to 1024 entries on the latest uarchs.

A cached TLB miss will take about 10-20 cycles (progressively more if the lookup
chain misses in the cache) - but that cost is partially hidden if the L1 data
cache was missed (which is likely for most TLB-flush intense workloads), and will
be almost completely hidden if it goes out to the L3 cache or goes to RAM. (It
takes up cache/memory bandwidth though, but unless the access patters are totally
sparse, it should be a small fraction.)

A single INVLPG with its 200+ cycles cost is equivalent to about 10-20 TLB misses.
That's a lot.

So this kind of workload should trigger the TLB flushing 'worst case': with say
512 dTLB entries you could see up to 5k-10k cycles of hidden/indirect cost, but
potentially parallelized with other misses going on with the same data accesses.

The current limit for INVLPG flushing is 33 entries: that's 10k-20k cycles max
with an INVLPG cost of 250 cycles - this could explain the results you got.

But the problem is: AFAICS you can only decrease the INVLPG count by decreasing
the batching size - the additional IPI costs will overwhelm any TLB preservation
benefits. So depending on the cost relationship between INVLPG, TLB miss cost and
IPI cost, it might not be possible to see a speedup even in the worst-case.

Thanks,

Ingo


\
 
 \ /
  Last update: 2015-06-10 11:01    [W:0.369 / U:0.368 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site