Great work,
+1 to merge it to all branches.
Can you post results of the same test for xmlrpc?
I wonder what is the slowdown/speedup between protocols.
On Mon, Nov 10, 2014 at 9:02 AM, Francesco Romani <fromani(a)redhat.com> wrote:
Hi everyone,
I was doing JSON-RPC investigation recently, running VDSM on RHEL6.{5,6},
and while some initial profile work, I discovered a (little) performance degradation
about pure JSONRPC coding/encoding, *only* on the above platforms. Here's why:
- the JSON codec module in the python stdlib is based on simplejson
- (thus) simplejson API is identical to stdlib's json module API
- python 2.6 JSON codec is (based on) simplejson 1.9
- python 2.7 JSON codec is (based on) simplejson 2.0.9, with significant speedups. See
https://docs.python.org/2/whatsnew/2.7.html#new-and-improved-modules and
https://bugs.python.org/issue4136
- RHEL6.x/Centos 6.x includes python 2.6, so here (and only here) JSON codec
is unnecessarily slower.
- RHEL6.x (surely) and CentOS 6.x (need to check, don't see why not) includes
anyway simplejson 2.0.9 as external package
So, it seems to me that we can get python 2.7 speed even on stable platform with
a five line patch:
=== cut here ===
diff --git a/lib/yajsonrpc/__init__.py b/lib/yajsonrpc/__init__.py
index b3fd590..6682fc3 100644
--- a/lib/yajsonrpc/__init__.py
+++ b/lib/yajsonrpc/__init__.py
@@ -12,7 +12,10 @@
# You should have received a copy of the GNU General Public
# License along with this program; if not, write to the Free Software
# Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
-import json
+try:
+ import simplejson as json
+except ImportError
+ import json
import logging
from functools import partial
from Queue import Queue
=== cut here ===
Here's an excerpt of the effects of this patch, using an initial still rough
bencmark
- once again 100 VMs boot, is the beginning of every testing routine from mine,
so nothing new here.
Let's consider two initial profiles. Please note that
1) the amount of calls is not the same (error from mine in profiling, need more precise
bench)
but still
2) using 2.0.9 module JSON just vanishes from the profile alltogether
Vanilla VDSM, stdlib json, top 20 expensive calls
ncalls tottime percall cumtime percall filename:lineno(function)
19/100 127.216 6.696 9.974 0.100
/usr/lib/python2.6/site-packages/mom/GuestMonitor.py:51(GuestMonitor.run)
1555/1635 41.180 0.026 58.566 0.036
/usr/lib64/python2.6/threading.py:481(Thread.run)
11708/1553160 11.967 0.001 37.921 0.000
/usr/lib64/python2.6/copy.py:144(deepcopy)
1268613 6.154 0.000 9.209 0.000
/usr/lib64/python2.6/copy.py:261(_keep_alive)
4890/130093 6.149 0.001 37.596 0.000
/usr/lib64/python2.6/copy.py:251(_deepcopy_dict)
497858/2070078 5.224 0.000 11.102 0.000
/usr/lib64/python2.6/json/encoder.py:284(JSONEncoder._iterencode)
498440/1333474 4.187 0.000 10.368 0.000
/usr/lib64/python2.6/json/encoder.py:213(JSONEncoder._iterencode_dict)
12/100 4.115 0.343 4.130 0.041
/usr/share/vdsm/virt/sampling.py:424(VmStatsThread.run)
43826 3.692 0.000 6.516 0.000
/usr/lib64/python2.6/xml/dom/expatbuilder.py:743(ExpatBuilderNS.start_element_handler)
7345 3.657 0.000 11.047 0.002
/usr/share/vdsm/virt/vm.py:2264(Vm._getRunningVmStats)
9666/320508 3.568 0.000 3.572 0.000
/usr/lib64/python2.6/xml/dom/minidom.py:305(_get_elements_by_tagName_helper)
274994/275000 3.339 0.000 5.593 0.000
/usr/lib64/python2.6/StringIO.py:208(StringIO.write)
90217 1.823 0.000 1.870 0.000
/usr/lib64/python2.6/xml/dom/minidom.py:349(Attr.__init__)
44025/44036 1.666 0.000 2.470 0.000
/usr/share/vdsm/storage/lvm.py:217(makeLV)
144212/144226 1.287 0.000 1.292 0.000
/usr/lib/python2.6/site-packages/vdsm/utils.py:285(convertToStr)
201 1.285 0.006 7.278 0.036
/usr/share/vdsm/storage/lvm.py:411(LVMCache._reloadlvs)
109880 1.235 0.000 1.283 0.000
/usr/lib64/python2.6/xml/dom/minidom.py:281(Document._append_child)
57874 1.163 0.000 1.799 0.000
/usr/lib64/python2.6/xml/dom/expatbuilder.py:274(ExpatBuilderNS.character_data_handler_cdata)
1130859 1.160 0.000 1.160 0.000
/usr/lib64/python2.6/copy.py:197(_deepcopy_atomic)
3021/3022 1.121 0.000 2.554 0.001
/usr/lib64/python2.6/inspect.py:247(getmembers)
patch applied, top 20 expensive calls
24/100 117.222 4.884 13.372 0.134
/usr/lib/python2.6/site-packages/mom/GuestMonitor.py:51(GuestMonitor.run)
1209/1302 60.812 0.050 64.963 0.050
/usr/lib64/python2.6/threading.py:481(Thread.run)
9243/1235834 10.259 0.001 33.047 0.000
/usr/lib64/python2.6/copy.py:144(deepcopy)
13/100 7.007 0.539 7.281 0.073
/usr/share/vdsm/virt/sampling.py:424(VmStatsThread.run)
4116/103669 5.536 0.001 32.779 0.000
/usr/lib64/python2.6/copy.py:251(_deepcopy_dict)
9670/320560 4.437 0.000 4.443 0.000
/usr/lib64/python2.6/xml/dom/minidom.py:305(_get_elements_by_tagName_helper)
5653 3.232 0.001 9.443 0.002
/usr/share/vdsm/virt/vm.py:2264(Vm._getRunningVmStats)
43836 3.117 0.000 7.304 0.000
/usr/lib64/python2.6/xml/dom/expatbuilder.py:743(ExpatBuilderNS.start_element_handler)
274967/275000 2.955 0.000 6.833 0.000
/usr/lib64/python2.6/StringIO.py:208(StringIO.write)
1100/28500 2.376 0.002 12.881 0.000
/usr/share/vdsm/virt/vmxml.py:489(Element.__hacked_writexml)
43745/44036 1.729 0.000 2.736 0.000
/usr/share/vdsm/storage/lvm.py:217(makeLV)
90217/90221 1.695 0.000 1.953 0.000
/usr/lib64/python2.6/xml/dom/minidom.py:349(Attr.__init__)
201 1.549 0.008 7.725 0.038
/usr/share/vdsm/storage/lvm.py:411(LVMCache._reloadlvs)
57888 1.441 0.000 2.216 0.000
/usr/lib64/python2.6/xml/dom/expatbuilder.py:274(ExpatBuilderNS.character_data_handler_cdata)
2351/2354 1.262 0.001 2.750 0.001
/usr/lib64/python2.6/inspect.py:247(getmembers)
992794 1.232 0.000 8.027 0.000
/usr/lib64/python2.6/copy.py:261(_keep_alive)
4827 1.150 0.000 1.368 0.000
/usr/lib/python2.6/site-packages/mom/Policy/spark.py:211(Parser.buildState)
108040 1.119 0.000 1.119 0.000
/usr/lib/python2.6/site-packages/vdsm/utils.py:285(convertToStr)
336771/422122 1.106 0.000 2.575 0.000
/usr/share/vdsm/protocoldetector.py:108(MultiProtocolAcceptor._process_events)
883021 0.987 0.000 0.987 0.000
/usr/lib64/python2.6/copy.py:197(_deepcopy_atomic)
Indeed this is a microbenchmark, but my point is we *recover* some speed at nearly
zero cost and risk.
Now, the question is:
Do we want this patch? Do we want it on master? On 3.5 branch? Only on RHEL/Centos 6
platforms?
--
Francesco Romani
RedHat Engineering Virtualization R & D
Phone: 8261328
IRC: fromani
_______________________________________________
Devel mailing list
Devel(a)ovirt.org
http://lists.ovirt.org/mailman/listinfo/devel