Zmq随机接收和回复

Question

I have two question regarding Zmq:-

1) I am implementing a server which would receive 10K requests/per minute and then redirect these request towards backend. After getting response from backend it will then send response back. I can't use request/reply as restriction in this pattern is recv/reply should be synchronous. Can anyone suggest me as what to use in this scenario.

2) Do I need to implement multi-threading in this also.

Thanks in advance.

Answer 1

Let me answer first the Q1 - requirement on performance & Q2 - if there is a need for multithreading:

What are ZeroMQ transport overhead times about?

From trivial performance tests, intentionally adding a TCP-transport wrapping overhead, you may taste the "cost" of ZeroMQ message handling layer in [usec] per a message sent.

                         21.936  [usec]/MSG
                        111.280  [usec]/MSG
                         39.714  [usec]/MSG
                         37.080  [usec]/MSG
                         11.351  [usec]/MSG

To have an idea about a message batch size impact relief, lets read retest with growin size

>>> [ sender( nBatchSIZE = x ) for x in ( 1, 10, 100, 1000, 10000, 100000, 1000000 ) ]

sent       1 ... RUN took:       58  [usec] i.e.  58.0  [usec]/MSG
sent      10 ... RUN took:      156  [usec] i.e.  15.6  [usec]/MSG
sent     100 ... RUN took:     1071  [usec] i.e.  10.7  [usec]/MSG
sent    1000 ... RUN took:    10561  [usec] i.e.  10.5  [usec]/MSG
sent   10000 ... RUN took:   106478  [usec] i.e.  10.6  [usec]/MSG
sent  100000 ... RUN took:  1333242  [usec] i.e.  13.3  [usec]/MSG

Based on these figures, your Q1 the system will have no problem with 10k/min stream of events on reasonable sized MSGs ( not mentioning multi-GB BLOBs )

The processing performance is thus limited by the back-end phase

Suggestion on Q2-what to use :

The REQ/REP thus goes out of the table. Rather than multithreading thus try to consider any performance scale-able approaches to increase your end-to-end processing capability -- using more advanced schemes for load-balancers alike REQ/ROUTER||ROUTER/REQ et al, you may both increase the processing speed and add some failure resolution features. Check sample trivialised structures in Single- & Multi-Cluster Architectures

Answer 2

For your message volume, it's all about message size. Assuming a roughly "typical" message size (say up to several hundred bytes would be typical) and a system that can otherwise handle that throughput (eg if you were to preload your messages rather than receive them with ZMQ, would it process your message volume efficiently enough?), then you won't break a sweat with 10k messages/minute. You could probably do that each second (as you can see in the other answer).

ZMQ will support whatever message pattern you're looking for, but you'll have to provide more information to determine what pattern would be optimal for you. There are a lot of good examples in the guide .