RE: [RFC 0/2] Add high-performance timer facility

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Mattias Rönnblom [mailto:mattias.ronnblom@ericsson.com]
> Sent: Tuesday, 28 February 2023 10.39

I have been looking for a high performance timer library (for use in a fast path TCP stack), and this looks very useful, Mattias.

My initial feedback is based on quickly skimming the patch source code, and reading this cover letter.

> 
> This patchset is an attempt to introduce a high-performance, highly
> scalable timer facility into DPDK.
> 
> More specifically, the goals for the htimer library are:
> 
> * Efficient handling of a handful up to hundreds of thousands of
>   concurrent timers.
> * Reduced overhead of adding and canceling timers.
> * Provide a service functionally equivalent to that of
>   <rte_timer.h>. API/ABI backward compatibility is secondary.
> 
> In the author's opinion, there are two main shortcomings with the
> current DPDK timer library (i.e., rte_timer.[ch]).
> 
> One is the synchronization overhead, where heavy-weight full-barrier
> type synchronization is used. rte_timer.c uses per-EAL/lcore skip
> lists, but any thread may add or cancel (or otherwise access) timers
> managed by another lcore (and thus resides in its timer skip list).
> 
> The other is an algorithmic shortcoming, with rte_timer.c's reliance
> on a skip list, which, seemingly, is less efficient than certain
> alternatives.
> 
> This patchset implements a hierarchical timer wheel (HWT, in

Typo: HWT or HTW?

> rte_htw.c), as per the Varghese and Lauck paper "Hashed and
> Hierarchical Timing Wheels: Data Structures for the Efficient
> Implementation of a Timer Facility". A HWT is a data structure
> purposely design for this task, and used by many operating system
> kernel timer facilities.
> 
> To further improve the solution described by Varghese and Lauck, a
> bitset is placed in front of each of the timer wheel in the HWT,
> reducing overhead of rte_htimer_mgr_manage() (i.e., progressing time
> and expiry processing).
> 
> Cycle-efficient scanning and manipulation of these bitsets are crucial
> for the HWT's performance.
> 
> The htimer module keeps a per-lcore (or per-registered EAL thread) HWT
> instance, much like rte_timer.c keeps a per-lcore skip list.
> 
> To avoid expensive synchronization overhead for thread-local timer
> management, the HWTs are accessed only from the "owning" thread.  Any
> interaction any other thread has with a particular lcore's timer
> wheel goes over a set of DPDK rings. A side-effect of this design is
> that all operations working toward a "remote" HWT must be
> asynchronous.
> 
> The <rte_htimer.h> API is available only to EAL threads and registered
> non-EAL threads.
> 
> The htimer API allows the application to supply the current time,
> useful in case it already has retrieved this for other purposes,
> saving the cost of a rdtsc instruction (or its equivalent).
> 
> Relative htimer does not retrieve a new time, but reuse the current
> time (as known via/at-the-time of the manage-call), again to shave off
> some cycles of overhead.

I have a comment to the two points above.

I agree that the application should supply the current time.

This should be the concept throughout the library. I don't understand why TSC is used in the library at all?

Please use a unit-less tick, and let the application decide what one tick means.

A unit-less tick will also let the application instantiate a HTW with higher resolution than the TSC. (E.g. think about oversampling in audio processing, or Brezenham's line drawing algorithm for 2D visuals - oversampling can sound and look better.)

For reference (supporting my suggestion), the dynamic timestamp field in the rte_mbuf structure is also defined as being unit-less. (I think NVIDIA implements it as nanoseconds, but that's an implementation specific choice.)

> 
> A semantic improvement compared to the <rte_timer.h> API is that the
> htimer library can give a definite answer on the question if the timer
> expiry callback was called, after a timer has been canceled.
> 
> Below is a performance data from DPDK's 'app/test' micro benchmarks,
> using 10k concurrent timers. The benchmarks (test_timer_perf.c and
> test_htimer_mgr_perf.c) aren't identical in their structure, but the
> numbers give some indication of the difference.
> 
> Use case               htimer  timer
> ------------------------------------
> Add timer                 28    253
> Cancel timer              10    412
> Async add (source lcore)  64
> Async add (target lcore)  13
> 
> (AMD 5900X CPU. Time in TSC.)
> 
> Prototype integration of the htimer library into real, timer-heavy,
> applications indicates that htimer may result in significant
> application-level performance gains.
> 
> The bitset implementation which the HWT implementation depends upon
> seemed generic-enough and potentially useful outside the world of
> HWTs, to justify being located in the EAL.
> 
> This patchset is very much an RFC, and the author is yet to form an
> opinion on many important issues.
> 
> * If deemed a suitable replacement, should the htimer replace the
>   current DPDK timer library in some particular (ABI-breaking)
>   release, or should it live side-by-side with the then-legacy
>   <rte_timer.h> API? A lot of things in and outside DPDK depend on
>   <rte_timer.h>, so coexistence may be required to facilitate a smooth
>   transition.

It's my immediate impression that they are totally different in both design philosophy and API.

Personal opinion: I would call it an entirely different library.

> 
> * Should the htimer and htw-related files be colocated with rte_timer.c
>   in the timer library?

Personal opinion: No. This is an entirely different library, and should live for itself in a directory of its own.

> 
> * Would it be useful for applications using asynchronous cancel to
>   have the option of having the timer callback run not only in case of
>   timer expiration, but also cancellation (on the target lcore)? The
>   timer cb signature would need to include an additional parameter in
>   that case.

If one thread cancels something in another thread, some synchronization between the threads is going to be required anyway. So we could reprase your question: Will the burden of the otherwise required synchronization between the two threads be significantly reduced if the library has the ability to run the callback on asynchronous cancel?

Is such a feature mostly "Must have" or "Nice to have"?

More thoughts in this area...

If adding and additional callback parameter, it could be an enum, so the callback could be expanded to support "timeout (a.k.a. timer fired)", "cancel" and more events we have not yet come up with, e.g. "early kick".

Here's an idea off the top of my head: An additional callback parameter has a (small) performance cost incurred with every timer fired (which is a very large multiplier). It might not be required. As an alternative to an "what happened" parameter to the callback, the callback could investigate the state of the object for which the timer fired, and draw its own conclusion on how to proceed. Obviously, this also has a performance cost, but perhaps the callback works on the object's state anyway, making this cost insignificant.

Here's another alternative to adding a "what happened" parameter to the callback:

The rte_htimer could have one more callback pointer, which (if set) will be called on cancellation of the timer.

> 
> * Should the rte_htimer be a nested struct, so the htw parts be separated
>   from the htimer parts?
> 
> * <rte_htimer.h> is kept separate from <rte_htimer_mgr.h>, so that
>   <rte_htw.h> may avoid a depedency to <rte_htimer_mgr.h>. Should it
>   be so?
> 
> * rte_htimer struct is only supposed to be used by the application to
>   give an indication of how much memory it needs to allocate, and is
>   its member are not supposed to be directly accessed (w/ the possible
>   exception of the owner_lcore_id field). Should there be a dummy
>   struct, or a #define RTE_HTIMER_MEMSIZE or a rte_htimer_get_memsize()
>   function instead, serving the same purpose? Better encapsulation,
>   but more inconvenient for applications. Run-time dynamic sizing
>   would force application-level dynamic allocations.
> 
> * Asynchronous cancellation is a little tricky to use for the
>   application (primarily due to timer memory reclamation/race
>   issues). Should this functionality be removed?
> 
> * Should rte_htimer_mgr_init() also retrieve the current time? If so,
>   there should to be a variant which allows the user to specify the
>   time (to match rte_htimer_mgr_manage_time()). One pitfall with the
>   current proposed API is an application calling rte_htimer_mgr_init()
>   and then immediately adding a timer with a relative timeout, in
>   which case the current absolute time used is 0, which might be a
>   surprise.
> 
> * Should libdivide (optionally) be used to avoid the div in the TSC ->
>   tick conversion? (Doesn't improve performance on Zen 3, but may
>   do on other CPUs.) Consider <rte_reciprocal.h> as well.
> 
> * Should the TSC-per-tick be rounded up to a power of 2, so shifts can be
>   used for conversion? Very minor performance gains to be found there,
>   at least on Zen 3 cores.
> 
> * Should it be possible to supply the time in rte_htimer_mgr_add()
>   and/or rte_htimer_mgr_manage_time() functions as ticks, rather than
>   as TSC? Should it be possible to also use nanoseconds?
>   rte_htimer_mgr_manage_time() would need a flags parameter in that
>   case.

Do not use TSC anywhere in this library. Let the application decide the meaning of a tick.

> 
> * Would the event timer adapter be best off using <rte_htw.h>
>   directly, or <rte_htimer.h>? In the latter case, there needs to be a
>   way to instantiate more HWTs (similar to the "alt" functions of
>   <rte_timer.h>)?
> 
> * Should the PERIODICAL flag (and the complexity it brings) be
>   removed? And leave the application with only single-shot timers, and
>   the option to re-add them in the timer callback.

First thought: Yes, keep it lean and remove the periodical stuff.

Second thought: This needs a more detailed analysis.

From one angle:

How many PERIODICAL versus ONESHOT timers do we expect?

Intuitively, I would use this library for ONESHOT timers, and perhaps implement my periodical timers by other means.

If the PERIODICAL:ONESHOT ratio is low, we can probably live with the extra cost of cancel+add for a few periodical timers.

From another angle:

What is the performance gain with the PERIODICAL flag?

Without a periodical timer, cancel+add costs 10+28 cycles. How many cycles would a "move" function, performing both cancel and add, use?

And then compare that to the cost (in cycles) of repeating a timer with PERIODICAL?

Furthermore, not having the PERIODICAL flag probably improves the performance for non-periodical timers. How many cycles could we gain here?


Another, vaguely related, idea:

The callback pointer might not need to be stored per rte_htimer, but could instead be common for the rte_htw.

When a timer fires, the callback probably needs to check/update the state of the object for which the timer fired anyway, so why not just let the application use that state to determine the appropriate action. This might provide some performance benefit.

It might complicate using one HTW for multiple different purposes, though. Probably a useless idea, but I wanted to share the idea anyway. It might trigger other, better ideas in the community.

> 
> * Should the async result codes and the sync cancel error codes be merged
>   into one set of result codes?
> 
> * Should the rte_htimer_mgr_async_add() have a flag which allow
>   buffering add request messages until rte_htimer_mgr_process() is
>   called? Or any manage function. Would reduce ring signaling overhead
>   (i.e., burst enqueue operations instead of single-element
>   enqueue). Could also be a rte_htimer_mgr_async_add_burst() function,
>   solving the same "problem" a different way. (The signature of such
>   a function would not be pretty.)
> 
> * Does the functionality provided by the rte_htimer_mgr_process()
>   function match its the use cases? Should there me a more clear
>   separation between expiry processing and asynchronous operation
>   processing?
> 
> * Should the patchset be split into more commits? If so, how?
> 
> Thanks to Erik Carrillo for his assistance.
> 
> Mattias Rönnblom (2):
>   eal: add bitset type
>   eal: add high-performance timer facility