Experimental Physics and
Industrial Control System

"Wang Xiaoqiang" <[email protected]> · Mon, 19 Sep 2011 16:39:46 +0200

Hi Matt,

In ca.create_channel method, after libca.ca_create_channel it does also a poll. It would be more efficient to poll after creating all the channels. 

I only have ~3000 channels to test. The connection time already shows some difference.
  Poll after each channel creation: 4.5 sec
  Poll after all channels creation: 0.35 sec

So Python program would in the end match the speed of sddscasr, whew..

BTW: I am doing this originally to see how CaChannel performs in comparison.

Best,
Xiaoqiang

-----Original Message-----
From: [email protected] [mailto:[email protected]] On Behalf Of Matt Newville
Sent: Saturday, September 17, 2011 5:40 PM
To: Matt Newville; EPICS tech-talk
Subject: Re: How to write a python-based backpup tool?

Hi Emmanuel,

OK, I also improved my earlier numbers with a more careful test.

With 20K PVs, all connected and all on a single subnet (but distributed over several IOCs, most of them on VME/vxWorks IOCs), I get about 13 to 14 sec total, or ~0.7usec / PV, with it basically breaking down as

  20K x (  pv = epics.PV(pname) ) :   10 to 11 sec
  20K x (  val = pv.get() )     :   2 to 3 sec
  write to disk    :   0.1 sec

I'm pretty sure most of that 10 to 11 sec for the PV "creation" is actually in the automated connection and event callbacks.

Using the lower-level epics.ca calls didn't change the total time at all. But that does still use a lot of python code, including implicit connection callback and unpacking of the values from ca_array_get(), etc.  I haven't tried using the simplest, most direct calls to the CA library.  I'm curious if there would be any improvement.

You say that sddscasr and burtrb are 5 times faster.  To be honest, I'd be surprised if a pyepics solution could be improved by a factor of 5, but it's worth looking into.

Anyway, my earlier estimates were definitely pessimistic (and my times probably dominated by unconnected PVs).

--Matt Newville <newville at cars.uchicago.edu> 630-252-0431

On Fri, Sep 16, 2011 at 5:38 PM,  <[email protected]> wrote:
>
>
> With same request files ( ~11K Pvs), I have sddscasr 2 sec burtrb   2 
> sec python  10 sec (with gigabit network)
>
> I am wondering what is the reason for the difference.
> As you can see below, just the connection to PV is comparatively 'slow'
>
> debug output:
> [Fri Sep 16 16:11:15 2011] import epics start [Fri Sep 16 16:11:15 
> 2011] import epics end [Fri Sep 16 16:11:15 2011] request read [Fri 
> Sep 16 16:11:21 2011] connected to pv [Fri Sep 16 16:11:23 2011] 
> fetched PV values [Fri Sep 16 16:11:23 2011] write on disk [Fri Sep 16 
> 16:11:23 2011] write completed
>
> with
>
> #!/usr/bin/env python
>
> WRITE_ON_DISK = True
> USE_SQLITE = False
>
> import time
>
> print "[%s] import epics start" % time.asctime( 
> time.localtime(time.time()) ) import epics print "[%s] import epics 
> end" % time.asctime( time.localtime(time.time()) )
>
>
> pvNames = []
> pvs = []
> vals = []
>
> with open("burt.req") as fp:
>    for line in iter(fp.readline, ''):
>        pvNames.append(line[:-1])
>
> print "[%s] request read" % time.asctime( time.localtime(time.time()) 
> )
>
> for pvName in pvNames:
>    pvs.append(epics.PV(pvName))
>
> print "[%s] connected to pv" % time.asctime( 
> time.localtime(time.time()) )
>
> for pv in pvs:
>    vals.append(pv.get())
>
> print "[%s] fetched PV values" % time.asctime( 
> time.localtime(time.time()) )
>
> if not WRITE_ON_DISK : exit()
>
> print "[%s] write on disk" % time.asctime( time.localtime(time.time()) 
> )
>
> if USE_SQLITE :
>    from pysqlite2 import dbapi2 as sqlite
>
>    connection = sqlite.connect('test.db')
>    cursor = connection.cursor()
>    cursor.execute('CREATE TABLE backup (id INTEGER PRIMARY KEY, pvname 
> VARCHAR(50), value VARCHAR(50))')
>
>    for index, pvName in enumerate(pvNames):
>        sqlquery = "INSERT INTO backup VALUES (null, '%s','%s')" % 
> (pvName, vals[index] )
>        #cursor.execute('INSERT INTO names VALUES (null, "%s" % pvName, 
> "%s" % vals[index])')
>        cursor.execute(sqlquery)
>
>
>    connection.commit()
>    cursor.close()
>    connection.close()
>
> else:
>    with open("test.txt",'w') as fp:
>        for index, pvName in enumerate(pvNames) :
>            print >>fp , "%s %s" % ( pvName, vals[index] )
>
>
> print "[%s] write completed" % time.asctime( 
> time.localtime(time.time()) )
>
>
>
>
>
>
> On 16:30 Fri 16 Sep     , Matt Newville wrote:
>> Hi Emmanuel,
>>
>> On Fri, Sep 16, 2011 at 3:09 PM,  <[email protected]> wrote:
>> >> pvlist = []
>> >> for pvName in pvNames:
>> >>     pvlist.append( epics.PV(pvName) )
>> >>
>> >> for pvs in pvlist:
>> >>     val = pv.get()
>> >>     <store value>
>> >>
>> >> With this sort of approach, I typically see on the order of 10ms 
>> >> per PV connection on startup.   That is, if I create and connect 
>> >> to 5K PVs, it takes ~50 seconds (meaning 40 to 80 seconds) to get 
>> >> initial values.  I believe that is mostly the CA library, not the 
>> >> python part, and I believe it would scale, suggesting that any 
>> >> save/restore process that runs once and then quit would take 10 minutes for 50K PVs.
>> >
>> > Are you saying that 10 min for 50K PVs is also what is observed with C utilities?
>> > 10 min for 50K is rather long ...
>>
>> I'd be happy to be shown wrong, and I'm certainly not the right 
>> person here to answer for sure,  but I think this might be the case.
>>
>> > If that is the case, then one whould be careful to backup only PV 
>> > which are of interest and not the 'entire' machine.
>> > How is this being managed at large installation?
>>
>> I think they break up the set of variables across 
>> processes/machines/NICs and/or have a long-running process that 
>> repeatedly saves variables instead of
>>     start process; connect to 50K PVs; write values; end process.
>>
>> > I understand your points, but somehow I am having a hard time believing this.
>> > Did anyone benchmark their backup platform?
>> > (I remember seeing a few powerpoints on this issue a while back...)
>>
>> For what it's worth, and also gladly proven wrong, etc:  I recall 
>> that with non-preemptive-callback context, creating and connecting 
>> PVs took exactly 30ms, to very high precision and reproducibility.  I 
>> never fully understood that....   From that view, the 10 ms average 
>> with the preemptive-context is an improvement.
>>
>> It's probably easy to tell whether it is closer to 1, 10, or 30 ms 
>> even with a few hundred PVs.  But even at 1ms / PV connection, you 
>> would probably want to not start a new connection per PV for saving 
>> 50K PVs.
>>
>> > PS:
>> > Is there a difference between a connection to 50K Pvs versus, 50K connections to the same PVs?
>>
>> Yes, these are definitely different.
>>
>> --Matt
>>
>
> --
> Emmanuel
>
>

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System