Re: [dpdk-announce] important design choices - statistics - ABI

["Dumitrescu, Cristian" <cristian.dumitrescu@intel.com>]

> -----Original Message-----
> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Matthew Hall
> Sent: Wednesday, June 17, 2015 5:37 AM
> To: dev@dpdk.org
> Subject: Re: [dpdk-dev] [dpdk-announce] important design choices -
> statistics - ABI
> 
> On Wed, Jun 17, 2015 at 01:29:47AM +0200, Thomas Monjalon wrote:
> > There were some debates about software statistics disabling.
> > Should they be always on or possibly disabled when compiled?
> > We need to take a decision shortly and discuss (or agree) this proposal:
> > 	http://dpdk.org/ml/archives/dev/2015-June/019461.html
> 
> This goes against the idea I have seen before that we should be moving
> toward
> a distro-friendly approach where one copy of DPDK can be used by multiple
> apps
> without having to rebuild it. It seems like it is also a bit ABI hostile
> according to the below goals / discussions.

Matthew, thanks for your input. As Thomas also mentions, it would be good to first read the proposal on the guideline (http://dpdk.org/ml/archives/dev/2015-June/019461.html).

In the guideline proposal, we are addressing the topic of preventing the ABI changes due to library statistics support, it would be good to get your opinion on that, I am not sure you saw it. Given DPDK paramount focus on performance, I think probably provides the best solution for ABI compatibility in this case.

> 
> Jemalloc is also very high-performance code and still manages to allow
> enabling and disabling statistics at runtime. Are we sure it's impossible for
> DPDK or just theorizing?

A performance improvement for one environment might still be an overkill for DPDK run-time environment.

Stats cannot be enabled/disabled at run-time without performance impact, as it typically requires testing a persistent flag for incrementing the stats counters, which requires a conditional branch instruction. Sometimes the branches are predicted correctly, sometimes not, and in this case the CPU pipeline needs to be flushed; typical branch misprediction cost is ~14 cycles, which is important given a packet budget of ~200 cycles.

When only a small number of branches are present in the (application + library) code, they are predicted correctly, so I am sure that for a simple test application like l3fw the cost of run-time enabled stats is not visible. The problem is that real applications are much more complex and they typically have a lot of conditional branches, out of which the library stats test is just one of them, so they cannot be all predicted correctly, hence the cost of stats run-time enablement becomes very much visible.

> 
> > During the development of the release 2.0, there was an agreement to
> keep
> > ABI compatibility or to bring new ABI while keeping old one during one
> release.
> > In case it's not possible to have this transition, the (exceptional) break
> > should be acknowledged by several developers.
> 
> Personally to me it seems more important to preserve the ABI on patch
> releases, like 2.X.Y going to 2.X.Z. But maybe I missed something?
> 
> > During the current development cycle for the release 2.1, the ABI question
> > arises many times in different threads.
> 
> Most but not all of these examples point to a different issue which
> sometimes
> happens in libraries... often seen as "old-style" versus "new-style" C library
> interface. For example, in old-style like libpcap there are a lot of structs,
> both opaque and non-opaque, which the caller must allocate in order to run
> libpcap.
> 
> However new-style libraries such as libcurl usually just have init functions
> which initialize all the secret structs based on some defaults and some user
> parameters and hide the actual structs from the user. If you want to adjust
> some you call an adjuster function that modifies the actual secret struct
> contents, with some enum saying what field to adjust, and the new value
> you
> want it to have.
> 
> If you want to keep a stable ABI for a non-stable library like DPDK, there's a
> good chance you must begin hiding all these weird device specific structs all
> over the DPDK from the user needing to directly allocate and modify them.
> Otherwise the ABI breaks everytime you have to add adjustments,
> extensions,
> modifications to all these obscure special features.
> 
> Matthew.