1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
|
==========
Benchmarks
==========
pyftpdlib 0.7.0 vs. pyftpdlib 1.0.0
-----------------------------------
+-----------------------------------------+-----------------+----------------+------------+
| *benchmark type* | *0.7.0* | *1.0.0* | *speedup* |
+=========================================+=================+================+============+
| STOR (client -> server) | 528.63 MB/sec | 585.90 MB/sec | **+0.1x** |
+-----------------------------------------+-----------------+----------------+------------+
| RETR (server -> client) | 1702.07 MB/sec | 1652.72 MB/sec | -0.02x |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (connect, login) | 1.70 secs | 0.19 secs | **+8x** |
+-----------------------------------------+-----------------+----------------+------------+
| STOR (1 file with 300 idle clients) | 60.77 MB/sec | 585.59 MB/sec | **+8.6x** |
+-----------------------------------------+-----------------+----------------+------------+
| RETR (1 file with 300 idle clients) | 63.46 MB/sec | 1497.58 MB/sec | **+22.5x** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (RETR 10M file) | 4.68 secs | 3.41 secs | **+0.3x** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (STOR 10M file) | 10.13 secs | 8.78 secs | **+0.1x** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (QUIT) | 0.02 secs | 0.02 secs | 0x |
+-----------------------------------------+-----------------+----------------+------------+
pyftpdlib vs. proftpd 1.3.4
---------------------------
+-----------------------------------------+-----------------+----------------+------------+
| *benchmark type* | *pyftpdlib* | *proftpd* | *speedup* |
+=========================================+=================+================+============+
| STOR (client -> server) | 585.90 MB/sec | 600.49 MB/sec | -0.02x |
+-----------------------------------------+-----------------+----------------+------------+
| RETR (server -> client) | 1652.72 MB/sec | 1524.05 MB/sec | **+0.08** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (connect, login) | 0.19 secs | 9.98 secs | **+51x** |
+-----------------------------------------+-----------------+----------------+------------+
| STOR (1 file with 300 idle clients) | 585.59 MB/sec | 518.55 MB/sec | **+0.1x** |
+-----------------------------------------+-----------------+----------------+------------+
| RETR (1 file with 300 idle clients) | 1497.58 MB/sec | 1478.19 MB/sec | 0x |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (RETR 10M file) | 3.41 secs | 3.60 secs | **+0.05x** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (STOR 10M file) | 8.60 secs | 11.56 secs | **+0.3x** |
+-----------------------------------------+-----------------+----------------+------------+
| 300 concurrent clients (QUIT) | 0.03 secs | 0.39 secs | **+12x** |
+-----------------------------------------+-----------------+----------------+------------+
pyftpdlib vs. vsftpd 2.3.5
--------------------------
+-----------------------------------------+----------------+----------------+-------------+
| *benchmark type* | *pyftpdlib* | *vsftpd* | *speedup* |
+=========================================+================+================+=============+
| STOR (client -> server) | 585.90 MB/sec | 611.73 MB/sec | -0.04x |
+-----------------------------------------+----------------+----------------+-------------+
| RETR (server -> client) | 1652.72 MB/sec | 1512.92 MB/sec | **+0.09** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (connect, login) | 0.19 secs | 20.39 secs | **+106x** |
+-----------------------------------------+----------------+----------------+-------------+
| STOR (1 file with 300 idle clients) | 585.59 MB/sec | 610.23 MB/sec | -0.04x |
+-----------------------------------------+----------------+----------------+-------------+
| RETR (1 file with 300 idle clients) | 1497.58 MB/sec | 1493.01 MB/sec | 0x |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (RETR 10M file) | 3.41 secs | 3.67 secs | **+0.07x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (STOR 10M file) | 8.60 secs | 9.82 secs | **+0.07x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (QUIT) | 0.03 secs | 0.01 secs | +0.14x |
+-----------------------------------------+----------------+----------------+-------------+
pyftpdlib vs. Twisted 12.3
--------------------------
By using *sendfile()* (Twisted *does not* support sendfile()):
+-----------------------------------------+----------------+----------------+-------------+
| *benchmark type* | *pyftpdlib* | *twisted* | *speedup* |
+=========================================+================+================+=============+
| STOR (client -> server) | 585.90 MB/sec | 496.44 MB/sec | **+0.01x** |
+-----------------------------------------+----------------+----------------+-------------+
| RETR (server -> client) | 1652.72 MB/sec | 283.24 MB/sec | **+4.8x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (connect, login) | 0.19 secs | 0.19 secs | +0x |
+-----------------------------------------+----------------+----------------+-------------+
| STOR (1 file with 300 idle clients) | 585.59 MB/sec | 506.55 MB/sec | **+0.16x** |
+-----------------------------------------+----------------+----------------+-------------+
| RETR (1 file with 300 idle clients) | 1497.58 MB/sec | 280.63 MB/sec | **+4.3x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (RETR 10M file) | 3.41 secs | 11.40 secs | **+2.3x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (STOR 10M file) | 8.60 secs | 9.22 secs | **+0.07x** |
+-----------------------------------------+----------------+----------------+-------------+
| 300 concurrent clients (QUIT) | 0.03 secs | 0.09 secs | **+2x** |
+-----------------------------------------+----------------+----------------+-------------+
By using plain *send()*:
+-----------------------------------------+----------------+---------------+--------------+
| *benchmark type* | *tpdlib* | *twisted* | *speedup* |
+=========================================+================+===============+==============+
| RETR (server -> client) | 894.29 MB/sec | 283.24 MB/sec | **+2.1x** |
+-----------------------------------------+----------------+---------------+--------------+
| RETR (1 file with 300 idle clients) | 900.98 MB/sec | 280.63 MB/sec | **+2.1x** |
+-----------------------------------------+----------------+---------------+--------------+
Memory usage
------------
*Values on UNIX are calculated as (rss - shared).*
+------------------------------------------+-------------+-----------------+----------------+----------------+
| *benchmark type* | *pyftpdlib* | *proftpd 1.3.4* | *vsftpd 2.3.5* | *twisted 12.3* |
+==========================================+=============+=================+================+================+
| Starting with | 6.7M | 1.4M | 352.0K | 13.4M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| STOR (1 client) | 6.7M | 8.5M | 816.0K | 13.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| RETR (1 client) | 6.8M | 8.5M | 816.0K | 13.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| 300 concurrent clients (connect, login) | **8.8M** | 568.6M | 140.9M | 13.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| STOR (1 file with 300 idle clients) | **8.8M** | 570.6M | 141.4M | 13.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| RETR (1 file with 300 idle clients) | **8.8M** | 570.6M | 141.4M | 13.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| 300 concurrent clients (RETR 10.0M file) | **10.8M** | 568.6M | 140.9M | 24.5M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
| 300 concurrent clients (STOR 10.0M file) | **12.6** | 568.7M | 140.9M | 24.7M |
+------------------------------------------+-------------+-----------------+----------------+----------------+
Interpreting the results
------------------------
pyftpdlib, `proftpd`_ and `vsftpd`_ look pretty much equally fast. The huge
difference is noticeable in scalability though, because of the concurrency
model adopted. Proftpd and vsftpd spawn a new process for every connected
client, whereas pyftpdlib doesn't (see `the C10k problem`_). The difference
can be noticed on connect/login benchmarks and memory benchmarks.
The huge differences between 0.7.0 and 1.0.0 versions of pyftpdlib are due to
fix of `issue 203`_ . On Linux we now use `epoll()`_ which scales considerably
better than `select()`_. The fact that we're downloading a file with 300 idle
clients doesn't make any difference for `epoll()`. We might as well had 5000
idle clients and the result would have been the same. On Windows, where we
still use select(), 1.0.0 still wins hands down as the asyncore loop was
reimplemented from scratch in order to support fd un/registration and
modification. Benchmarks were conducted on Linux Ubuntu 12.04, Intel core duo -
3.1 Ghz box.
Setup
-----
The following setup was used before running every benchmark:
proftpd config
^^^^^^^^^^^^^^
::
# /etc/proftpd/proftpd.conf
MaxInstances 2000
...followed by:
::
$ sudo service proftpd restart
vsftpd config
^^^^^^^^^^^^^
::
# /etc/vsftpd.conf
local_enable=YES
write_enable=YES
max_clients=2000
max_per_ip=2000
...followed by:
::
$ sudo service vsftpd restart
twisted FTP server
^^^^^^^^^^^^^^^^^^
::
from twisted.protocols.ftp import FTPFactory, FTPRealm
from twisted.cred.portal import Portal
from twisted.cred.checkers import AllowAnonymousAccess, FilePasswordDB
from twisted.internet import reactor
import resource
soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
resource.setrlimit(resource.RLIMIT_NOFILE, (hard, hard))
open('pass.dat', 'w').write('user:some-passwd')
p = Portal(FTPRealm('./'),
[AllowAnonymousAccess(), FilePasswordDB("pass.dat")])
f = FTPFactory(p)
reactor.listenTCP(21, f)
reactor.run()
...followed by:
::
$ sudo python3 twist_ftpd.py
pyftpdlib
^^^^^^^^^
The following patch was applied first:
::
Index: pyftpdlib/servers.py
===================================================================
--- pyftpdlib/servers.py (revisione 1154)
+++ pyftpdlib/servers.py (copia locale)
@@ -494,3 +494,10 @@
def _map_len(self):
return len(multiprocessing.active_children())
+
+import resource
+soft, hard = resource.getrlimit(resource.RLIMIT_NOFILE)
+resource.setrlimit(resource.RLIMIT_NOFILE, (hard, hard))
+FTPServer.max_cons = 0
...followed by:
::
$ sudo python3 demo/unix_daemon.py
The `benchmark script`_ was run as:
::
python3 scripts/ftpbench -u USERNAME -p PASSWORD -b all -n 300
...and for the memory test:
::
python3 scripts/ftpbench -u USERNAME -p PASSWORD -b all -n 300 -k FTP_SERVER_PID
.. _`benchmark script`: https://github.com/giampaolo/pyftpdlib/blob/master/scripts/ftpbench
.. _`epoll()`: https://linux.die.net/man/4/epoll
.. _`issue 203`: https://github.com/giampaolo/pyftpdlib/issues/203
.. _`proftpd`: http://www.proftpd.org/
.. _`select()`: https://linux.die.net/man/2/select
.. _`the C10k problem`: http://www.kegel.com/c10k.html
.. _`vsftpd`: https://security.appspot.com/vsftpd.html
|
