Josh,
Thanks for the detailed response. I would like to avoid going through a
separate process if at all possible due to the performance penalty. I have
also tried your last suggestion to create a dedicated pytables thread and
send everything through that but still see the same problem (Runtime error
in flush). This leads me to believe something strange is going on behind
the scenes. ??
Updated test program with dedicated pytables thread reading an input
Queue.Queue:
import tables
import threading
import random
import time
import Queue
# a simple table
class TableValue(tables.IsDescription):
a = tables.Int64Col(pos=1)
b = tables.UInt32Col(pos=2)
class TablesThread(threading.Thread):
def __init__(self):
threading.Thread.__init__(self)
self.name = 'HDF5 io thread'
# create the dummy HDF5 file
self.h5 = None
self.h5 = tables.openFile('/data/test.h5', mode='w')
self.num_groups = 5
self.num_tables = 5
self.groups = [self.h5.createGroup('/', "group%d"%i) for i in
range(self.num_groups)]
self.tables = []
for group in self.groups:
tbls = [self.h5.createTable(group, 'table%d'%i, TableValue) for
i in range(self.num_tables)]
self.tables.append (tbls)
for table in tbls:
# add an index for good measure
table.cols.a.createIndex()
self.stopEvt = threading.Event()
self.stoppedEvt = threading.Event()
self.inputQ = Queue.Queue()
def run(self):
try:
while not self.stopEvt.is_set():
# get a command
try:
cmd, args, result = self.inputQ.get(timeout = 0.5)
except Queue.Empty:
# poll stopEvt so we can shutdown
continue
# do the command
if cmd == 'write':
x = self.tables[args[0]][args[1]].row
x['a'] = args[2]
x['b'] = args[3]
x.append()
elif cmd == 'read':
self.h5.flush()
table = self.tables[args[0]][args[1]]
result.value = table.readWhere('a > %d'%(args[2]))
else:
raise Exception("Command not supported: %s"%(cmd,))
# signal that the result is ready
result.event.set()
finally:
# shutdown
self.h5.close()
self.stoppedEvt.set()
def stop(self):
if not self.stoppedEvt.is_set():
self.stopEvt.set()
self.stoppedEvt.wait()
class ResultEvent():
def __init__(self):
self.event = threading.Event()
self.value = None
class Test():
def __init__(self):
self.tables = TablesThread()
self.tables.start()
self.timeout = 5
self.stats = {'read': 0,
'write': 0,
'read_error': 0,
'write_error': 0}
def write(self):
r = ResultEvent()
self.tables.inputQ.put(('write',
(random.randint(0,
self.tables.num_groups-1),
random.randint(0,
self.tables.num_tables-1),
random.randint(0, 100),
random.randint(0, 100)),
r))
r.event.wait(timeout = self.timeout)
if r.event.is_set():
self.stats['write'] += 1
else:
self.stats['write_error'] += 1
def read(self):
r = ResultEvent()
self.tables.inputQ.put(('read',
(random.randint(0,
self.tables.num_groups-1),
random.randint(0,
self.tables.num_tables-1),
random.randint(0, 100)),
r))
r.event.wait(timeout = self.timeout)
if r.event.is_set():
self.stats['read'] += 1
#print 'Query got %d hits'%(len(r.value))
else:
self.stats['read_error'] += 1
def close(self):
self.tables.stop()
def __del__(self):
self.close()
class Worker(threading.Thread):
def __init__(self, method):
threading.Thread.__init__(self)
self.method = method
self.stopEvt = threading.Event()
def run(self):
while not self.stopEvt.is_set():
try:
self.method()
except Exception, x:
print 'Worker thread failed with: %s'%(x,)
time.sleep(random.random()/100.0)
def stop(self):
self.stopEvt.set()
def main():
t = Test()
threads = [Worker(t.write) for _i in range(10)]
threads.extend([Worker(t.read) for _i in range(10)])
for thread in threads:
thread.start()
time.sleep(5)
for thread in threads:
thread.stop()
for thread in threads:
thread.join()
t.close()
print t.stats
if __name__ == "__main__":
main()
On Wed, Dec 5, 2012 at 10:52 PM, Josh Ayers <josh.ay...@gmail.com> wrote:
> Alan,
>
> Unfortunately, the underlying HDF5 library isn't thread-safe by default.
> It can be built in a thread-safe mode that serializes all API calls, but
> still doesn't allow actual parallel access to the disk. See [1] for more
> details. Here's [2] another interesting discussion concerning whether
> multithreaded access is actually beneficial for an I/O limited library like
> HDF5. Ultimately, if one thread can read at the disk's maximum transfer
> rate, then multiple threads don't provide any benefit.
>
> Beyond the limitations of HDF5, PyTables also maintains global state in
> various module-level variables. One example is the _open_file cache in the
> file.py module. I made an attempt in the past to work around this to allow
> read-only access from multiple threads, but didn't make much progress.
>
> In general, I think your best bet is to serialize all access through a
> single process. There is another example in the PyTables/examples
> directory that benchmarks different methods of transferring data from
> PyTables to another process [3]. It compares Python's
> multiprocessing.Queue, sockets, and memory-mapped files. In my testing,
> the latter two are 5-10x faster than using a queue.
>
> Another option would be to use multiple threads, but handle all access to
> the HDF5 file in one thread. PyTables will release the GIL when making
> HDF5 library calls, so the other threads will be able to run. You could
> use a Queue.Queue or some other mechanism to transfer data between
> threads. No actual copying would be needed since their memory is shared,
> which should make it faster than the multi-process techniques.
>
> Hope that helps.
>
> Josh Ayers
>
>
> [1]: http://www.hdfgroup.org/hdf5-quest.html#mthread
>
> [2]:
> https://visitbugs.ornl.gov/projects/8/wiki/Multi-threaded_cores_and_HPC-HDF5
>
> [3]:
> https://github.com/PyTables/PyTables/blob/develop/examples/multiprocess_access_benchmarks.py
>
>
> On Wed, Dec 5, 2012 at 2:24 PM, Alan Marchiori <a...@alanmarian.com>wrote:
>
>> I am trying to allow multiple threads read/write access to pytables data
>> and found it is necessary to call flush() before any read. If not, the
>> latest data is not returned. However, this can cause a RuntimeError. I
>> have tried protecting pytables access with both locks and queues as done by
>> joshayers (
>> https://github.com/PyTables/PyTables/blob/develop/examples/multiprocess_access_queues.py).
>> In either case I still get RuntimeError: dictionary changed size during
>> iteration when doing the flush. (incidentally using the Locks appears to
>> be much faster than using queues in my unscientific tests...)
>>
>> I have tried versions 2.4 and 2.3.1 with the same results. Interestingly
>> this only appears to happen if there are multiple tables/groups in the H5
>> file. To investigate this behavior further I create a test program to
>> illustrate (below). When run with num_groups = 5 num_tables = 5 (or
>> greater) I see the runtime error every time. When these values are smaller
>> than this it doesn't (at least in a short test period).
>>
>> I might be doing something unexpected with pytables, but this seems
>> pretty straight forward to me. Any help is appreciated.
>>
>>
>>
------------------------------------------------------------------------------
LogMeIn Rescue: Anywhere, Anytime Remote support for IT. Free Trial
Remotely access PCs and mobile devices and provide instant support
Improve your efficiency, and focus on delivering more value-add services
Discover what IT Professionals Know. Rescue delivers
http://p.sf.net/sfu/logmein_12329d2d
_______________________________________________
Pytables-users mailing list
Pytables-users@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/pytables-users
