WSGI POST reading

[skinkie]

I am currently trying to benchmark wsgi apps with posting. What puzzles me is that it seems that every framework I try POSTing creates a delay of one second on top of the processing time.

If I compare fapws3 with uwsgi, the POST performance is similar, where uwsgi outperforms fapws3 if a client posts to the server, while the server is not touching the wsgi.input object. When enabling post buffering in uwsgi, a similar latency of fapws3 is observed.

Reproducability:

import fapws._evwsgi as evwsgi
from fapws import base

def KV55(environ, start_response):
    #response = environ['wsgi.input'].read(int(environ['CONTENT_LENGTH']))
    response = ''
    start_response('200 OK', [('Content-Type', 'text/xml'), ('Content-length', str(len(response)))])
    yield response

def start():
    evwsgi.start("0.0.0.0", "8181")
    evwsgi.set_base_module(base)

    evwsgi.wsgi_cb(("/", KV55))

    evwsgi.set_debug(0)
    evwsgi.run()

start()

time curl -X POST -T kv55.xml http://127.0.0.1:8181

real 0m1.014s user 0m0.008s sys 0m0.004s

(document is about 3k)

Because I started with uwsgi, I figured out that the delay seems to be on an epoll_wait.

{{EPOLLIN, {u32=3, u64=4294967299}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19638, 918400814}) = 0 accept(3, {sa_family=AF_INET, sin_port=htons(48250), sin_addr=inet_addr("127.0.0.1")}, [16]) = 6 fcntl(6, F_GETFL) = 0x2 (flags O_RDWR) fcntl(6, F_SETFL, O_RDWR|O_NONBLOCK) = 0 epoll_ctl(4, EPOLL_CTL_ADD, 6, {EPOLLIN, {u32=6, u64=12884901894}}) = 0 clock_gettime(CLOCK_MONOTONIC, {19638, 918913760}) = 0 epoll_wait(4, {{EPOLLIN, {u32=6, u64=12884901894}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19638, 919034276}) = 0 read(6, "POST /kv55.xml HTTP/1.1\r\nUser-Ag"..., 32768) = 196 clock_gettime(CLOCK_MONOTONIC, {19638, 919240711}) = 0

epoll_wait(4, {{EPOLLIN, {u32=6, u64=12884901894}}}, 64, 59743) = 1

clock_gettime(CLOCK_MONOTONIC, {19639, 919742520}) = 0 read(6, "<?xml version=\"1.0\"?>\n\n"..., 32768) = 3935 epoll_ctl(4, EPOLL_CTL_MOD, 6, {EPOLLOUT, {u32=6, u64=17179869190}}) = 0 clock_gettime(CLOCK_MONOTONIC, {19639, 920070711}) = 0 epoll_wait(4, {{EPOLLOUT, {u32=6, u64=17179869190}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19639, 920205867}) = 0 write(6, "HTTP/1.0 200 OK\r\nDate: Tue, 24 J"..., 119) = 119 clock_gettime(CLOCK_MONOTONIC, {19639, 920982086}) = 0 epoll_wait(4, {{EPOLLOUT, {u32=6, u64=17179869190}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19639, 921086163}) = 0 close(6) = 0 clock_gettime(CLOCK_MONOTONIC, {19639, 921306357}) = 0

[jonashaag]

How about code tags. Works like this

 ```py
 your code here

[skinkie]

Anyway small update. Roberto De Ioris gave me the insight that this is because of the Expect header. Applying -H "Expect:" to curl, gives instant performance again.

[william-os4y]

Hmmm interesting. This sounds more linked yo libev it self.

But what i would like to change the post processing. Indeed currently fapws stores the data in memory. This could be a big problem if the user send a very big file. We should have a kind of temporary storage on disk. And we should provide the possibility to the developer to block too big upload.

Regards

W. On Jan 24, 2012 3:26 AM, "Stefan de Konink" < reply@reply.github.com> wrote:

I am currently trying to benchmark wsgi apps with posting. What puzzles me is that it seems that every framework I try POSTing creates a delay of one second on top of the processing time.

If I compare fapws3 with uwsgi, the POST performance is similar, where uwsgi outperforms fapws3 if a client posts to the server, while the server is not touching the wsgi.input object. When enabling post buffering in uwsgi, a similar latency of fapws3 is observed.

Reproducability: import fapws._evwsgi as evwsgi from fapws import base

def KV55(environ, start_response):

response = environ['wsgi.input'].read(int(environ['CONTENT_LENGTH']))

response = '' start_response('200 OK', [('Content-Type', 'text/xml'), ('Content-length', str(len(response)))]) yield response

def start(): evwsgi.start("0.0.0.0", "8181") evwsgi.set_base_module(base)

evwsgi.wsgi_cb(("/", KV55))

evwsgi.set_debug(0) evwsgi.run()

start()

time curl -X POST -T kv55.xml http://127.0.0.1:8181

real 0m1.014s user 0m0.008s sys 0m0.004s

(document is about 3k)

Because I started with uwsgi, I figured out that the delay seems to be on an epoll_wait.

{{EPOLLIN, {u32=3, u64=4294967299}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19638, 918400814}) = 0 accept(3, {sa_family=AF_INET, sin_port=htons(48250), sin_addr=inet_addr("127.0.0.1")}, [16]) = 6 fcntl(6, F_GETFL) = 0x2 (flags O_RDWR) fcntl(6, F_SETFL, O_RDWR|O_NONBLOCK) = 0 epoll_ctl(4, EPOLL_CTL_ADD, 6, {EPOLLIN, {u32=6, u64=12884901894}}) = 0 clock_gettime(CLOCK_MONOTONIC, {19638, 918913760}) = 0 epoll_wait(4, {{EPOLLIN, {u32=6, u64=12884901894}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19638, 919034276}) = 0 read(6, "POST /kv55.xml HTTP/1.1\r\nUser-Ag"..., 32768) = 196 clock_gettime(CLOCK_MONOTONIC, {19638, 919240711}) = 0

epoll_wait(4, {{EPOLLIN, {u32=6, u64=12884901894}}}, 64, 59743) = 1

clock_gettime(CLOCK_MONOTONIC, {19639, 919742520}) = 0 read(6, "<?xml version=\"1.0\"?>\n\n"..., 32768) = 3935 epoll_ctl(4, EPOLL_CTL_MOD, 6, {EPOLLOUT, {u32=6, u64=17179869190}}) = 0 clock_gettime(CLOCK_MONOTONIC, {19639, 920070711}) = 0 epoll_wait(4, {{EPOLLOUT, {u32=6, u64=17179869190}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19639, 920205867}) = 0 write(6, "HTTP/1.0 200 OK\r\nDate: Tue, 24 J"..., 119) = 119 clock_gettime(CLOCK_MONOTONIC, {19639, 920982086}) = 0 epoll_wait(4, {{EPOLLOUT, {u32=6, u64=17179869190}}}, 64, 59743) = 1 clock_gettime(CLOCK_MONOTONIC, {19639, 921086163}) = 0 close(6) = 0 clock_gettime(CLOCK_MONOTONIC, {19639, 921306357}) = 0

---

Reply to this email directly or view it on GitHub: https://github.com/william-os4y/fapws3/issues/32

[william-os4y]

I don't know exactly what you add in the header with the -H of curl, but, please consider the simple following POST script: "import socket

f = open('test', 'r') #this could be replaced by your xml file strf = f.read() f.close() s = socket.socket(socket.AF_INET, socket.SOCK_STREAM) s.connect(('127.0.0.1', 8181)) post_data = "POST / HTTP/1.1\r\nContent-Length: %s\r\n\r\n%s\r\n" % (len(strf),strf) print(post_data) s.send(post_data)

print("RECV:",s.recv(1024))

s.close() "

wi@myarchhost:/tmp$ time python2.7 send.py real 0m0.042s user 0m0.027s sys 0m0.010s

I've tried with epoll and select (LIBEV_FLAGS), in both cases I received the same response time.

I close the issue.

[skinkie]

curl -X POST -T kv55.xml http://127.0.0.1:8181

vs

curl -X POST -H "Expect:" -T kv55.xml http://127.0.0.1:8181