update settings faq for GPU vs CPU only

doc/Frequently-asked-questions.md

@@ -11,19 +11,45 @@ Exploiting the `503` mechanism is already implemented in the different GROBID cl
 
 ## Could we have some guidance for server configuration in production?
 
-The exact server configuration will depend on the service you want to call. We present here the configuration used to process with `processFulltextDocument` around 10.6 PDF per second (around 915,000 PDF per day, around 20M pages per day) with the node.js client listed above during one week on a 16 CPU machine (16 threads, 32GB RAM, no SDD). It ran without any crash during 7 days at this rate. We processed 11.3M PDF in a bit less than 7 days with two 16-CPU servers like that in one of our projects. 
+The exact server configuration will depend on the service you want to call, the models selected in the Grobid configuration file (`grobid-home/config/grobid.yaml`) and the availability of GPU. We consider here the complete full text processing of PDF (`processFulltextDocument`). 
 
-- if your server has 8-10 threads available, you can use the default settings of the docker image, otherwise you would rather need to build and start the service yourself to tune the parameters
+1) Using CRF models only, for example via the lightweight Docker image (https://hub.docker.com/r/lfoppiano/grobid/tags) 
 
-- keep the concurrency at the client (number of simultaneous calls) slightly higher than the available number of threads at the server side, for instance if the server has 16 threads, use a concurrency between 20 and 24 (it's the option `n` in the above mentioned clients, in my case I used 24)
+- in `grobid/grobid-home/config/grobid.yaml` set the parameter `concurrency` to your number of available threads at server side or slightly higher (e.g. 16 to 20 for a 16 threads-machine)
 
-- in `grobid/grobid-home/config/grobid.yaml` set the parameter `concurrency` to your number of available thread at server side or slightly higher (e.g. 16 to 20 for a 16 threads-machine, in my case I used 20)
+- keep the concurrency at the client (number of simultaneous calls) slightly higher than the available number of threads at the server side, for instance if the server has 16 threads, use a concurrency between 20 and 24 (it's the option `n` in the above mentioned clients)
 
-- set `modelPreload` to `true`in `grobid/grobid-home/config/grobid.yaml`, it will avoid some strange behavior at launch 
+These settings will ensure that CPU are fully used when processing a large set of PDF.  
 
-- in the query, `consolidateHeader` can be `1`  or `2` if you are using the biblio-glutton or CrossRef consolidation. It significantly improves the accuracy and add useful metadata.
+For example, with these settings, we processed with `processFulltextDocument` around 10.6 PDF per second (around 915,000 PDF per day, around 20M pages per day) with the node.js client during one week on a 16 CPU machine (16 threads, 32GB RAM, no SDD). It ran without any crash during 7 days at this rate. We processed 11.3M PDF in a bit less than 7 days with two 16-CPU servers in one of our projects. 
 
-- If you want to consolidate all the bibliographical references and use `consolidateCitations` as `1` or `2`, CrossRef query rate limit will avoid scale to more than 1 document per second... For scaling the bibliographical reference resolution, you will need to use a local consolidation service, [biblio-glutton](https://github.com/kermitt2/biblio-glutton). The overall capacity will depend on the biblio-glutton service then, and the number of elasticsearch nodes you can exploit. From experience, it is difficult to go beyond 300K PDF per day when using consolidation for every extracted bibliographical references. 
+Note: if your server has 8-10 threads available, you can use the default settings of the docker image, otherwise you will need to modify the configuration file to tune the parameters, as [documented](Configuration.md).
+
+2) Using Deep Learning models, for example via the full Docker image (<https://hub.docker.com/r/grobid/grobid/tags>) 
+
+2.1) If the server has a GPU
+
+In case the server has a GPU, which has its own memory, the Deep Learning inferences are automatically parallelized on this GPU, without impacting the CPU and RAM memmory. The settings given above in 1) can normally be use similarly.
+
+2.2) If the server has CPU only
+
+When Deep Learning models run as well on CPU as fallback, the CPU are used more intensively (DL models push CPU computations quite a lot), more irregularly (Deep Learning models are called at certain point in the overall process, but not continuously) and the CPU will use additional RAM memory to load those larger models. For the DL inference on CPU, an additional thread is created, allocating its own memory. We can have up to 2 times more CPU used at peaks, and approx. up to 50% more memory. 
+
+The settings should thus be considered as follow: 
+
+- in `grobid/grobid-home/config/grobid.yaml` set the parameter `concurrency` to your number of available threads at server side divided by 2 (8 threads available, set concurrency to `4`)
+
+- keep the concurrency at the client (number of simultaneous calls) at the same level as the `concurrency` parameter at server side, for instance if the server has 16 threads, use a `concurrency` of `8` and the client concurrency at `8` (it's the option `n` in the clients)
+
+In addition, consider more RAM memory when running Deep Learning model on CPU, e.g. 24-32GB memory with concurrency at `8` instead of 16GB.
+
+3) In general, consider also these settings:
+
+- Set `modelPreload` to `true` in `grobid/grobid-home/config/grobid.yaml`, it will avoid some strange behavior at launch (this is the default setting).
+
+- Regarding the query parameters, `consolidateHeader` can be `1`  or `2` if you are using the biblio-glutton or CrossRef consolidation. It significantly improves the accuracy and add useful metadata.
+
+- If you want to consolidate all the bibliographical references and use `consolidateCitations` as `1` or `2`, the CrossRef query rate limit will make the scaling to more than 1 document per second impossible (so Grobid would typically wait 90% or more of its time waiting for CrossRef API responses)... For scaling the bibliographical reference resolutions, you will need to use a local consolidation service, [biblio-glutton](https://github.com/kermitt2/biblio-glutton). The overall capacity will depend on the biblio-glutton service then, and the number of elasticsearch nodes you can exploit. From experience, it is difficult to go beyond 300K PDF per day when using consolidation for every extracted bibliographical references. 
 
 ## I would also like to extract images from PDFs