Hi Chris, Thanks for the clarification and the 'minimal spec'. I'll try to hack something together for that.
Where does the 'value' in your example come from? J Von meinem Mobiltelefon gesendet -------- Ursprüngliche Nachricht -------- Betreff: Re: [CodeGen] Performance values Von: Christofer Dutz An: [email protected] Cc: Actually, the operations you are concerned about are the only ones we don't need ;-) The parsing and serialization has some expressions in the field definitions. Currently they are: payload.lengthInBytes + 4 lengthInBytes - (payload.lengthInBytes + 1) (headerLength + 1) - curPos Rest parameter.lengthInBytes payload.lengthInBytes items.size() numItems address.lengthInBytes lengthInBytes == 8 parameter.numItems so you see it's not that complicated ... and I actually don't want to run any expression evaluator, but I want a parser so I can used the expression language ast to transform these expressions to something that can be run natively: lengthInBytes - (payload.lengthInBytes + 1) would be translated into the following Java code: value.getLengthInBytes() - (value.getPayload().getLengthInBytes + 1) JEXL is already quite potent and fast, but still using a runtime for interpreting this stuff at runtime costs a lot more than native code. Chris Am 10.06.19, 13:52 schrieb "Julian Feinauer" <[email protected]>: Hi Lukasz, good point. In fact, one reason why I initially favored to define these types in some kind of programming Language (python?!) was to avoid the creation of the N+1-th expression language. And those two where only examples. In facto you are right that we need to implement "special" operators (">>", "<<") for our scenarios. But nonetheless, it’s a pain to bother users with another language that does everything a bit cleverer than everyone else before (tm) but I see no way around it. Or has anyone else an idea? One idea could be to use a "known" language like Javascript or python... but that’s also not perfect, I think. Julian PS.: There are nice ways to make an IntelliJ language plugin as well as VS Code plugins from ANTLR grammars, this is something we should definetly address when we are settled. Am 10.06.19, 13:44 schrieb "Łukasz Dywicki" <[email protected]>: Keep in mind that Camel's "simple language" is N-th incarnation of same thing. As far I remember initial version of it was mixture (or rather torture) of reflection API combined with bunch of if-then-else statements. It was refactored to proper expression language between versions. In case of drivers we have lots of bit operations which are not even expressed in SpEL or Camel. We are lucky that syntax (or rules) for these are already known and used elsewhere. Both Spring and Camel work at "object" level thus we might reuse their experiences for higher level, where actual frame is constructed and we have linked parts such your examples with size. A difficult part will come with nested structures since none of Camel or SpEL works with these, yet frames can be enveloped multiple times. We need to retain an open way to let data and expressions pass through transport/network (TPDU/NPDU), protocol and application layers (APDU). If we leave an API/SPI for expression evaluation then anyone can pick a defined expression, transform it to own language, and let it be evaluated at runtime. Cheers, Łukasz -- ConnectorIO http://connectorio.com On 10.06.2019 11:43, Julian Feinauer wrote: > Hi Chris, > > I would really really really (...) love to jump in... but I'm a bit stressed out at the moment. > My approach would be to have a well specified language (similar to Spring EL, or Camels Simple Language) which is "well documented" and parsed via ANTLR (4!) to one of my ASTs because I can then generate "bodies" in arbitrary languages. > This should take about one day, I think, so perhaps I can contribute something soon. > > Perhaps it would be good to specify what we want to support and what the "default" methods are... like "getSize()" or so. > > Julian > > Am 10.06.19, 11:38 schrieb "Christofer Dutz" <[email protected]>: > > Hi all, > > While thinking of it, I remembered that abut 10 years ago I once created exactly such a grammar, but as ANTLR3 version. > So what I'll do, is create a new grammar which is aimed only as Expression language cause I think this could be something useful in general. > Perhaps you all could give some feedback as soon as I've got something. > > Chris > > Am 10.06.19, 10:40 schrieb "Christofer Dutz" <[email protected]>: > > Hi all, > > so on my trip home a few days ago I managed to also get the serialization running. I now am able to parse a byte message > into a model and deserialize the model back to bytes and the byte arrays are equal. However the serialization performance > I am not that happy with as it takes quite a lot longer to serialize than to parse, which shouldn't be the case. > > The main reason is, that while simply reading the implicit fields during the parsing, when serializing them, a lot of > Evaluations executions have to be performed. > They are usually quite simple expressions such as this: > exItems = jexl.createExpression("parameter.numItems"); > > The best option would be to improve the antlr grammar to parse the expressions a little more formally correct and to implement a model for these expressions and have them automatically translated to code like: > this.getParameter().getNumItems(); > > It should be possible and a lot faster ... anyone up for the challenge? @Julian? .. could you please help with this? > As you did that great job with the initial spec ANTLR grammar. > > Chris > > > > Am 05.06.19, 10:09 schrieb "Christofer Dutz" <[email protected]>: > > Hi all, > > In the train today I'll be working on the serialization (Which will be a challenge) > But I am sure this will be a lot of hard work but also a great step forward. > > Is there any progress on the Driver-Logic generation front? > > Otherwise I would probably try to whip up a hand-written Netty layer using the generated model. > Without all the parser/serializer code this should only be a fragment of the existing driver code. > > Chris > > > > > Am 05.06.19, 09:59 schrieb "Strljic, Matthias Milan" <[email protected]>: > > Hura sounds nice 😉 > I hope I find time to play a bit around with it in the next few days. > > Greetings > Matthias Strljic, M.Sc. > > Universität Stuttgart > Institut für Steuerungstechnik der Werkzeugmaschinen und Fertigungseinrichtungen (ISW) > > Seidenstraße 36 > 70174 Stuttgart > GERMANY > > Tel: +49 711 685-84530 > Fax: +49 711 685-74530 > > E-Mail: [email protected] > Web: http://www.isw.uni-stuttgart.de > > -----Original Message----- > From: Christofer Dutz <[email protected]> > Sent: Tuesday, June 4, 2019 4:16 PM > To: [email protected] > Subject: [CodeGen] Performance values > > Hi all, > > so as I mentioned in Slack yesterday I was able to successfully parse a S7 packet with the code generated by the code-generator. > There I am using Apache Jexl for evaluating the expressions we are using all over the place. It got things working quite easily. > However my gut-feeling told me all these Jexl evaluators I’m creating can’t be that ideal. > > But not wanting to pre-maturely optimize something that’s already good, I did a measurement: > > So I did a little test, in which I let my parser parse one message 20000 times. > > It came up with an average time of 0,8ms … this didn’t sound too bad compared to the about 20ms of the interpreted daffodil approach. > But is this already good? It’s probably not ideal to compare some results with the ones we know are bad, instead I wanted to compare it to the ones we are proud of. > > In order to find out I whipped up a little manual test with the existing S7 driver. > For this I simply plugged the 3 layers together with an Embedded channel and used a custom handler at the end to return the message. > This seems to work quite nicely and I was able to run the same test with the Netty based S7 driver layers we have been using for the last 2 years. > > The results were: > Parsed 20000 packets in 796ms > That's 0.0398ms per packet > > So this is a HUGE difference . > > As one last check I ran JProfiler over the benchmark and it confirmed that 87% of the time was used by Jexl. > However the creation of Jexl expressions, not their evaluation. > So one optimization I’ll try is to do, is to have the expressions created statically and then to simply reference them. > This will increase the complexity of the template, but should speed things up. > And I’ll also change the code I’m generating for the Type-Switches to work without jexl. > > Simply assuming this would eliminate the time wasted by jexl (great simplification), we would reduce the parse time to 0,1ms which is still about 3 times that of the existing driver. > I am assuming that this might be related to the BitInputStream I am using … but we’ll deal with that as soon as we’re done with getting rid of the time wasted by Jexl. > > So far an update on the generated-drivers front. > > Chris > > > > > > > > > >
