ZeroRPC performance. What is wrong with my Benchmark?

Question

I wrote a simple script to try to figure out how does ZeroRPC perform in terms of throughput of messages, the server is a simple service, echoing a greeting. The code of the client is below, see that I'm trying to launch parallel tasks:

import zerorpc
import datetime
import time
import gevent
import threading

N_MSGS      = 1000
N_TASKS     =   10
N_TASK_STEP =    1

count = dict()

client = zerorpc.Client()
client.connect('tcp://192.168.144.142:80081')

def task(number):
    results = []
    for i in range(N_MSGS):
        results.append(client.hello('Mathias', async=True))
        gevent.sleep(0)

    count[number] = 0
    for r in results:
        if r.get() == 'Hello Mathias':
            count[number] += 1


format_header = '{:<6s} {:<8s} {:<20s} {:<20s} {:<20s}'
format_line   = '{:>6d} {:>8d} {:>20s} {:>20d} {:>9.2f}'
print(format_header.format("#RUN", "#TASKS", "TOTAL TIME", "#MSGS", "MSG/SEC"))

run = 1
for i in range(1, N_TASKS, N_TASK_STEP):
    tasks=list()
    for j in range(i):
        tasks.append(gevent.spawn(task, j))

    start = datetime.datetime.now()
    gevent.joinall(tasks)
    end = datetime.datetime.now()

    total_time = end - start

    count_total = 0
    for _,v in count.items():
        count_total += v

    msg_per_sec = count_total / total_time.total_seconds()


    print(format_line.format(run, i, str(total_time),
                             count_total, msg_per_sec))
    run += 1

The server is quite simple:

import zerorpc

class Hello:
    def hello(self, name):
        return 'Hello {}'.format(name)

server = zerorpc.Server(Hello())
server.bind('ipc:///tmp/local')
server.run()

note that I'm using the IPC socket, I was expecting that whenever more tasks are added, the throughput would increase until a certain cap. But actually the numbers are almost the same always. Even when I change to a real network setup, the numbers do not differ a lot. Follow the numbers I got with the setup of the code above:

N_RUN   N_TASKS   TOTAL_TIME           N_MSGS               MSG/SEC              // vvvvvvv???????????
    1         1   0:00:01.543374         2000               1295.86              // 1295.8621824651705
    2         2   0:00:02.454205         4000               1629.85              // 1629.8556966512579
    3         3   0:00:03.583786         6000               1674.20              // 1674.2071094646835
    4         4   0:00:04.903248         8000               1631.57              // 1631.5715623602969
    5         5   0:00:06.133924        10000               1630.27              // 1630.2777797703395
    6         6   0:00:07.299903        12000               1643.85              // 1643.8574594758315
    7         7   0:00:10.096884        14000               1386.56              // 1386.5663901853286
    8         8   0:00:10.437927        16000               1532.87              // 1532.8714216912995
    9         9   0:00:11.384918        18000               1581.03              // 1581.0390553537582
   10        10   0:00:12.628328        20000               1583.74              // 1583.740935458756 
   11        11   0:00:13.691057        22000               1606.88              // 1606.888350548829 
   12        12   0:00:15.430392        24000               1555.37              // 1555.372021657    
   13        13   0:00:16.775109        26000               1549.91              // 1549.9154133663155
   14        14   0:00:18.021384        28000               1553.70              // 1553.7097483744865

I also took some measurements with IPC, and also I was expecting a bigger difference in performance when compared to TCP.

I plotted a chart, see below:

I have the impression that something is limiting the performance from increasing further.

Maybe I'm missing something in my code, I'm considering using ZeroRPC in an important project and would like to have a glimpse on its performance.

Am I doing something wrong in my benchmark?
Is this level of performance expected?
Could there be any limit on the number of connections in the client?
If yes could I increase the number?

Update <<<<

Just to be clear, since there is some information missing. I did run the parallel task, for the script above, and when I start two client process, the throughput drops by half, let's say one single client process it gets 2000 if I start a second I get 1000 for each. I'm not an expert on performance, I just decide to test by myself, since I see the framework as a nice potential for a project where I'm working on.

ZeroRPC performance. What is wrong with my Benchmark?

Update <<<<

Answers (1)

Related Questions