Hello there, Wanted to follow up on this. You didn't describe the setup, but I presume that all of your client threads are sending on the same channel to the server - is that correct? gRPC multiplexes streams (RPCs) onto a client channel, but ultimately the kernel-level polling activities are on the level of a FD (TCP connection). As a result, multiple client threads will need activity on one particular FD and one of them will be the one to get it. That thread is then responsible for processing the TCP activity that allows it to unblock the thread that is actually waiting for the particular stream. So, the more threads you have, the greater the chance will be that the thread that wakes up because of the OS activity is not the one that actually is waiting for the particular incoming stream. Additionally, don't forget to consider queueing latency at the clients: even if each client RPC is processed by a different thread, the shared socket is only serviced by one at a time.
As a side note, I am interested in your use of MM1 queueing theory to describe this issue. My gut feeling is that that's not the right model since I don't expect a memoryless service time ever, and the server consists of numerous parts: serialized socket processing but fully parallel RPC handling. Are you developing (or have you developed) a paper or TR validating this concept? Best regards - vjpai On Friday, September 29, 2017 at 10:23:07 AM UTC-7, [email protected] wrote: > > > Hello, > > I am using grpc for a university research project. > My experiment setup is as follows: I have a synchronous server running on > one machine. On another machine, I start a process that creates and > launches multiple threads (say X threads) that act as clients, and make > synchronous rpc calls to this synchronous server. The client threads are > then in the blocked state, waiting for responses from the server. > I want to find out how each client thread gets woken up when the > synchronous rpc call completes, and the response is returned to the client. > Is it that grpc has its own underlying thread pool where threads are woken > up at random until the correct thread is woken up, which then notifies the > corresponding thread in the 'X' client threads that I've created? > > I ask because I notice a strange behavior: As I go on increasing the > number of client threads, the responses take longer and longer to return to > each client. Using MM1 queuing theory, I eliminated queuing at the server > as the reason: the actual response latencies are orders of magnitude > greater than what is contributed by queuing. Also, upon looking at time > stamps on the send request to server and receive response from server > paths, I find that the receive response from server path is the culprit. > This possibly implies that the client is not able to pick up the response > exactly when the response is received, even though it was blocked waiting > for this response. So, my hypothesis is that random threads are woken up, > until the correct one is identified, which then waked up the client thread > that I created. > -- You received this message because you are subscribed to the Google Groups "grpc.io" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at https://groups.google.com/group/grpc-io. To view this discussion on the web visit https://groups.google.com/d/msgid/grpc-io/12b7ef79-ab5b-4cc9-808d-ccf534f01130%40googlegroups.com. For more options, visit https://groups.google.com/d/optout.
