Re: [PATCH RFC] introduce dl_iterate_phdr_parallel

[Gleb Natapov <gleb at scylladb dot com>]

On Thu, Aug 04, 2016 at 07:16:55PM +0200, Torvald Riegel wrote:
> On Wed, 2016-08-03 at 20:47 +0300, Gleb Natapov wrote:
> > On Wed, Aug 03, 2016 at 07:26:26PM +0200, Torvald Riegel wrote:
> > > On Mon, 2016-08-01 at 23:41 +0300, Gleb Natapov wrote:
> > > > On Mon, Aug 01, 2016 at 10:19:55PM +0200, Torvald Riegel wrote:
> > > > > On Mon, 2016-08-01 at 23:07 +0300, Gleb Natapov wrote:
> > > > > > On Mon, Aug 01, 2016 at 09:46:35PM +0200, Torvald Riegel wrote:
> > > > > > > If we built something custom for this and are willing to make the
> > > > > > > wrlock / exclusive-access case much more costly, we can decrease this
> > > > > > > overhead.  This could be roughly similar to one lock per thread or a set
> > > > > > > of rwlocks as you mentioned, but with less space overhead.
> > > > > > > 
> > > > > > IMO space overhead is negligible. More efficient rwlock is, of course,
> > > > > > better and can be useful in many more places. If you have something to
> > > > > > test I am willing to do so, but if custom rwlock will take time to
> > > > > > materialize we may start with lock array and change it later. The lock
> > > > > > array is not part of the interface, but implementation detail. What I
> > > > > > would like to avoid is stalling the afford while waiting for something
> > > > > > better. Exception scalability is very pressing issue for us.
> > > > > 
> > > > > I haven't looked at the workload in detail, and so can't say off-hand
> > > > > what would be best and how long it would take.  Is there a summary
> > > > > describing the synchronization that is necessary?  Once I know that, I
> > > > > can give a closer estimate.  It could be simple to do, so it might be
> > > > > worth having a look before trying the (rw?)lock-array approach.
> > > > > 
> > > > rwlock semantics is pretty much what is needed. There is a list of
> > > > loaded objects that changes rarely (only during dlopen/dlclose), but it
> > > > is accessed for read on each exception.
> > > 
> > > Does "for read" mean that the only thing you are interested in is some
> > > consistent snapshot of the list, or something like a yes/no result for a
> > > query about whether some element is contained?
> > No. The list contains objects that have pointers into some other memory that
> > the locking of the list prevents from been unmapped.
> 
> Then maybe something more specific to this case like hazard pointers
> would be better.  This could be used for other things in glibc too, and
> you wouldn't need atomic RMW as for the lock (which still have more
> overhead then stores, even though the RMWs are not as costly anymore on,
> for example, x86 as in the past).
> 
Access is done outside of glibc control. User code does regular
dereference.

> > > Additionally, the problem with the lock array is that you have to put
> > > the locks somewhere, preferably so that they are in a page that is local
> > > to where the thread runs. This makes a simple lock array suboptimal, at
> > > least on bigger systems.  TLS space would be most likely to be backed by
> > > local memory, but as Adhemerval has mentioned, we might have lots of
> > > threads, and if we do TLS, we need to sync between thread destruction
> > > and the heavyweight operation that wants to acquire all locks.
> > > 
> > The lock array is global, no TLS is involved. There is no problem of a
> > kind you describe.
> 
> If the lock array is global, it will be backed by memory to a single
> NUMA node.  This is inefficient if you access it from a CPU on a
> different NUMA node in a larger system.  You could split the array
> between NUMA nodes, but we don't have a mechanism to access CPU-local
> memory efficiently right now (it can be approximated, or one can try to
> amortize over several calls, but that's it).
> So the problem I am describing is having to make a choice between this
> problem and the problems that come with putting the lock into TLS.
> 
The NUMA problem is not different from having just one lock as we do
now. Can we do something better, yes. Having per thread lock for instance,
but I do not see a point in more complex solution without evidence of
real performance issues in proposed one.

> > > > New rwlock is substantially slower on only 4
> > > > cores (it is much better than current rwlock since it avoids entering
> > > > kernel if there are only readers).
> > > 
> > > Is this a microbenchmark just for acquisition/release of the rwlock, or
> > > does this significant difference also show up in the real scenario you
> > > are concerned about (the exception storm you mentioned)?
> > That's on a micro benchmark. But it simulates pretty closely what
> > sometimes happen on our real system when multiple threads encounter
> > burst of exceptional conditions (network connection dropped and all
> > queued requests, sometimes thousands, terminate with exception).
> 
> So the micro benchmark does the exception stuff in the rwlock critical
> sections (eg, you throw exceptions from multiple threads concurrently
> and measure overall throughput)?
What do you mean by "does the exception stuff in the rwlock critical
sections"? Stack unwinding happens in rwlock critical sections, the
benchmark just throws exception concurrently.

> (I'm asking so that I know whether the twice-as-slow claim you made
> (IIRC) really holds in the real-world use case too.)
For works case scenario it will. For best case even current code works
fine.

> 
> > > > and the data accessed
> > > > varies for each user of dl_iterate_phdr (elements of the list point into
> > > > mmaped parts of .so files).
> > > 
> > > That doesn't matter.  The only overhead is how often we need to validate
> > > so that we know that a subsequent pointer dereference is safe.  The
> > > number of these valuations compared to a lock acquisition+release is the
> > > deciding factor here.
> > Pointer dereferencing happens outside of glibc in a user provided
> > callback.
> 
> Ah, right.  I must have skipped the last part of the last sentence you
> wrote previously.
> 
> Is there a way to make callbacks only work on memory that is not going
> to be unmapped (e.g., under the assumption that the user won't ever
> dlclose stuff that's currently needed for any exception processing)?
> 
Callback may work on any memory of a loaded object. That's the point of
the interface - to run some code on each loaded object. What may work is
an RCU that will delay object deletion until quiescent point, but there
us no such points in glibc. I do not see how RCU can be implemented in glibc
unfortunately.

> (Note that one reason I'm asking these questions is to collect the
> information we need to document why we picked on or the other option to
> synchronize this.  Documenting, for example, "Gleb says seqlock don't
> work" isn't sufficient because it doesn't tell future glibc developers
> why that would be so.)
> 
Sure.

--
			Gleb.