Re: [dpdk-dev] [PATCH v2 00/16] update and simplify telemetrylibrary.

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, Apr 10, 2020 at 08:06:23PM +0200, Morten Brørup wrote:
> > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Wiles, Keith
> > Sent: Friday, April 10, 2020 4:22 PM
> > 
> > > On Apr 10, 2020, at 5:49 AM, Morten Brørup <mb@smartsharesystems.com>
> > wrote:
> > >
> > >> From: Ciara Power [mailto:ciara.power@intel.com]
> > >> Sent: Wednesday, April 8, 2020 6:50 PM
> > >>
> > >> This patchset extensively reworks the telemetry library adding new
> > >> functionality and simplifying much of the existing code, while
> > >> maintaining backward compatibility.
> > >>
> > >> This work is based on the previously sent RFC for a "process info"
> > >> library: https://patchwork.dpdk.org/project/dpdk/list/?series=7741
> > >> However, rather than creating a new library, this patchset takes
> > >> that work and merges it into the existing telemetry library, as
> > >> mentioned above.
> > >>
> > >> The telemetry library as shipped in 19.11 is based upon the metrics
> > >> library and outputs all statistics based on that as a source. However,
> > >> this limits the telemetry output to only port-level statistics
> > >> information, rather than allowing it to be used as a general scheme for
> > >> telemetry information across all DPDK libraries.
> > >>
> > >> With this patchset applied, rather than the telemetry library being
> > >> responsible for pulling ethdev stats and pushing them into the metrics
> > >> library for retrieval later, each library e.g. ethdev, rawdev, and even
> > >> the metrics library itself (for backwards compatiblity) now handle
> > >> their
> > >> own stats.  Any library or app can register a callback function with
> > >> telemetry, which will be called if requested by the client connected
> > >> via
> > >> the telemetry socket.
> > >
> > > Great. Standardization across libraries is a good improvement.
> > >
> > >> The callback function in the library/app then
> > >> formats its stats, or other data, into a JSON string, and returns it to
> > >> telemetry to be sent to the client.
> > >
> > > I am strongly opposed to using JSON as the standard format in DPDK, and
> > instead prefer a binary format with zero (or minimal) type conversion.
> > >
> > > Here is one reason why I dislike JSON for this: A part of our application
> > samples 100k+ counters every second to be able to provide drill-down
> > statistics through the GUI. Converting these counters from uint64_t to JSON
> > and back to uint64_t for data processing is not going to fly. And I assume
> > that we are not the only company developing equipment with drill-down
> > statistics.
> > >
> > > I am aware that there is a difference between statistics for *drill-down
> > and data processing* purposes and statistics for *telemetry eyeball viewing
> > only* purposes, but the line is blurry, and I see a big risk of setting a
> > path that leads to JSON being used in places where it shouldn't.
> > >
> > > Here is another reason why I dislike JSON for this: JSON is fine for the
> > LAMP stack with REST protocols. But other systems use other protocols with
> > other formats, e.g. the TICK stack uses an even simpler text based format.
> > So DPDK based systems supporting the TICK stack will need to convert to
> > first JSON format (in the DPDK libraries), and then from JSON format to
> > InfluxDB format (in the DPDK application).
> > >
> > > I think that type conversion does not belong inside deep inside the DPDK
> > libraries, but is a job for the DPDK application. However, DPDK could
> > provide libraries for efficient bulk conversion to popular formats like
> > JSON. And other formats, if they are relevant, e.g. ASN.1 used by old
> > school SNMP.
> > 
> > I believe JSON has it place in this library and in DPDK as it is a good
> > conversion tool and easy to utilize with a huge number of tools/languages.
> 
> JSON is extremely heavy compared to a raw binary format.
> 
> It makes sense for low volume, hierarchical structured data, but not for large tables or arrays of counters.
> 
> > Binary output gets into endianness issues and a number of other problems,
> > so I would not want all of the data exported from DPDK to be in binary
> > format.
> 
> Endianness considerations are only relevant for data exchanged across the network; not data exchanged across processes inside the same machine.
> 
> And if you are exchanging data across the network, you would usually implement one or more well known protocols for that, e.g. JSON over HTTPS, or ASN.1 over SNMP, or InfluxDB over UDP. This means that the application needs to implement a protocol handler, which - in my opinion - should handle the relevant data type conversions from the raw format provided by DPDK.
> 
> I think it would be silly for DPDK core libraries to provide counters in JSON format, so an SNMP Agent would need to convert them from JSON back to binary and then to ASN.1.
> 
> > If the layout of the structure changes then the code would need to
> > know that on both side to be able to convert the data into the correct
> > values.
> 
> I may be exaggerating here, but trying to prove a point: This is what we have ABI stability for. Structures should be designed cleverly and future proof, e.g. like the ethdev xstats. Using text based APIs is a circumvention of ABI stability.
> 
> > 
> > With that stated, the new telemetry code allows the application to add new
> > commands and with that you can create a binary set of commands along side
> > the JSON or any other output format. With the new register command we can
> > create say a ‘/ethdevraw/stats,X’ set of commands that can emit binary
> > format.
> 
> That would be silly. The protocol handler should make the protocol specific conversion, not the driver! Again, going to the extreme to prove a point: If I understand you correctly, this would mean that PMDs would have provide counters in ASN.1 format for SNMP.
> 
> Our application provides a HTTPS/REST based communication interface for multiple purposes, e.g. getting tables of data. And if you want to get a table of some data via this interface, you can specify the output format in the request, so you can get it in e.g. TSV format (tabulator separated with a headline) for scripts, HTML format for human eyeballs. This data conversion happens at a common location, so we can easily add other output formats. You don't want to push this all the way down to the originator of the data.
> 
> > 
> > Using this method we get the best of both worlds and when using languages
> > like Go or Python to collect these stats we have a standard format for
> > conversion. In Go it is pretty hard to do binary conversion and JSON
> > conversion is just a few lines. 
> 
> I don't think DPDK should provide preferential treatment to Go or Python. DPDK is based on C, and should mainly cater for C.
> 
> > JSON may not be the fastest, but if you are
> > requesting stats faster than a second then use the raw commands to get the
> > data, which anyone can add to its application or we can add them to DPDK as
> > a standard command set.
> 
> APIs in the libraries are currently available to get data in raw format. My main concern is that libraries in the future will not provide functions to get raw data, like they do now, but only JSON formatted data for the telemetry library. This is what I want to avoid.
> 

Hi Morten,

thanks for all the feedback.

Firstly on the performance side, we did some basic benchmarking of the
output capabilities of the current proposal. Using a dummy client that had
two queries constantly outstanding (to avoid dead time between one response
and next request), we measured the number of replies per second from the
ethdev xstats call.
* with a 2.3 GHz Xeon system, we got 3,500 responses per second, with 146
  stats per response, giving a total of >500k stats per second encoded and
  transmitted.
* for those 500k stats, looking at the cpu time on the core doing the
  telemetry work, ~80% of CPU time was spent inside the ixgbe driver
  getting the stats from the NIC card itself.
* replacing the ixgbe NIC with a ring vdev increased the responses per
  second to >100k, though with only 13 stats per response this time, giving
  1.3M stats per second.
* Splitting the existing xstats call in this RFC into separate xstat
  names and xstat values calls (something I think is a good idea to do
  generally), and only calling the xstat values call each time, gives
  further perf increases to 163k responses per second.

Overall so, at least on the json generation side, it appears we can do
things rather quickly, though I admit that we did not look into the parsing
cost on the other side.

All that being said, however, I do understand your point about having
everything work internally in json - something that ties in with Thomas
concern about having flexibility. Therefore, while we will upstream a v3
very shortly with the few incremental comments on v2, we can thereafter
look to switch the internal communication between callbacks and telemetry
library to use a regular data structure, and then leave the json encoding
to the library itself just before output. That would:
a) allow the addition of other output types in the future without needing
   new callbacks.
b) allow more flexibility for integrating with a future one-IPC mechanism
   for DPDK.

I hope this option will resolve most concerns and allow for a 20.05
integration.

Regards.
/Bruce