[BIOSAL] Causality and ACTION_STOP

Boisvert, Sebastien boisvert at anl.gov
Wed Nov 26 12:41:54 CST 2014 

Pulled.

[boisvert at bigmem biosal]$ git log --oneline master..gkt/granularity
becf1e6 hello_acq: clean up unneeded debug messages/make output a bit prettier
7e2c130 hello_acq: Revise code to have clear causality chain (tested on mpirun -np 1..16)
7b86f8e Merge branch 'master' of github.com:sebhtml/biosal into granularity
5a7b033 examples: update distributed hello to initiate greetings from only one actor


> From: George K. Thiruvathukal [gkt at cs.luc.edu]
> Sent: Wednesday, November 26, 2014 12:23 PM
> To: Boisvert, Sebastien
> Cc: biosal at lists.cels.anl.gov
> Subject: RE: Causality and ACTION_STOP
> 
> 
> 
> Seb, you can pull from granularity now. Do you need a formal pull request? I believe you are also OK with me mentioning to do same on the list.
> I'm going to work on the paper starting today as I was (finally) able to accomplish something with my hello code.
> I am becoming better at Thorium, though it will be awhile before I'm a true master like sebhtml. Luckily, I am getting smarter whenever I pull from sebhtml master. Sorry for the pun.
> On Nov 26, 2014 11:58 AM, "Boisvert, Sebastien" <boisvert at anl.gov> wrote:
> > From: George K. Thiruvathukal [gkt at cs.luc.edu]
> > Sent: Wednesday, November 26, 2014 11:50 AM
> > To: Boisvert, Sebastien
> > Cc: 
> biosal at lists.cels.anl.gov
> > Subject: Re: Causality and ACTION_STOP
> >
> >
> >
> >
> > Ok, great, so after a bit of thinking, I'm going to leave my version of hello_acquaintances as is and start working on something new. core_vector is actually a good match for what I'm trying to accomplish in that program. I would still end up having to iterate
> >  the keys of the core_map, which wouldn't be faster than working with a vector.
> 
> You should use "struct core_set" (basically like core_map, but only keys) for that.
> 
> The set could contain items that need attention. Once processed, remote the item.
> 
> You can get an item that needs processing (any of them) in O(1) with core_set_iterator,
> and delete is also O(1).
> 
> Can you send a message on the list for your pull request if you want me to pull ?
> 
> >
> >
> > I'm going to plan on working on the ideas I suggested as part of the more interesting example I'm trying to put together, which will also take advantage of actor spawning, etc.
> >
> >
> > I think I will use whatever cycles I have today to spend some time on the paper. :)
> >
> 
> Yes, the paper needs attention.
> 
> >
> > George
> >
> >
> >
> >
> >
> >
> >
> >
> > George K. Thiruvathukal, PhD
> >
> > Professor of Computer Science, Loyola University Chicago
> >
> > Director, Center for Textual Studies and Digital Humanities
> > Guest Faculty, Argonne National Laboratory, Math and Computer Science Division
> > Editor in Chief, Computing in
> >  Science and Engineering (IEEE CS/AIP)
> >
> > (w)
> > 
> thiruvathukal.com (v) 773.829.4872
> >
> >
> >
> >
> >
> >
> >
> >
> > On Wed, Nov 26, 2014 at 11:33 AM, Boisvert, Sebastien
> > <boisvert at anl.gov> wrote:
> >
> > You can use any action specifier as long as the source and the destination agree on it.
> >
> > You could use, for instance, ACTION_SUPER_COOL.
> >
> > > From: George K. Thiruvathukal [gkt at cs.luc.edu]
> > > Sent: Wednesday, November 26, 2014 11:26 AM
> >
> > > To: Boisvert, Sebastien
> > > Cc:
> > 
> biosal at lists.cels.anl.gov
> > > Subject: Re: Causality and ACTION_STOP
> > >
> > >
> > >
> > >
> > > Ok, I think I am mostly understanding the general idea. I may need you to walk me through this in person next week.
> > >
> > >
> > > So I should have an ACTION_PUSH_DATA and then I can just refer to the same variable (in my case, actors_to_greet) w/o doing an explicit unpack?
> > >
> > >
> > >
> > >
> > > George K. Thiruvathukal, PhD
> > >
> > > Professor of Computer Science, Loyola University Chicago
> > >
> > > Director, Center for Textual Studies and Digital Humanities
> > > Guest Faculty, Argonne National Laboratory, Math and Computer Science Division
> > > Editor in Chief, Computing in
> > >  Science and Engineering (IEEE CS/AIP)
> > >
> > > (w)
> > >
> > 
> thiruvathukal.com (v) 773.829.4872
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > >
> > > On Wed, Nov 26, 2014 at 11:14 AM, Boisvert, Sebastien
> > > <boisvert at anl.gov> wrote:
> > >
> > > > From: George K. Thiruvathukal [gkt at cs.luc.edu]
> > > > Sent: Wednesday, November 26, 2014 10:56 AM
> > > > To: Boisvert, Sebastien
> > > > Cc:
> > >
> > 
> biosal at lists.cels.anl.gov
> > > > Subject: Re: Causality and ACTION_STOP
> > > >
> > > >
> > > >
> > > >
> > > > Seb,
> > > >
> > > >
> > > > In your assembly code, I cannot find where the map that you packed is being unpacked. For completeness, I looked for all references to core_map_unpack in all of biosal:
> > > >
> > > >
> > > > mininuevo:biosal gkt$ grep core_map_unpack $(find . -name '*.[hc]')
> > > > ./core/structures/map.c:int core_map_unpack(struct core_map *self, void *buffer)
> > > > ./core/structures/map.h:int core_map_unpack(struct core_map *self, void *buffer);
> > > > ./genomics/data/coverage_distribution.c:        core_map_unpack(&map, buffer);
> > >
> > > You got it. coverage_distribution.c contains the script biosal_coverage_distribution.
> > >
> > >
> > > core_map_unpack is called when receiving ACTION_PUSH_DATA.
> > >
> > >
> > >
> > > On the sender side:
> > >
> > > Parallel code path: the message is sent here.
> > > assembly_graph_store.c 503
> > >     thorium_actor_send(self, customer, &new_message)  (with action specifier ACTION_PUSH_DATA,)
> > >
> > > customer is assigned here:
> > > assembly_graph_store.c 434
> > >     customer = concrete_self->customer;
> > >
> > > concrete_self->customer is set when receiving ACTION_SET_CONSUMER:
> > > assembly_graph_store.c 285
> > >         concrete_self->customer = customer;
> > >
> > >
> > >
> > >
> > >
> > > >
> > > > ./tests/test_map.c:        core_map_unpack(&map2, buffer);
> > > >
> > > > I can always use the example in test_map.c, but I'm trying to understand how your Spate code is unpacking the coverage_distribution map
> > >
> > > The sender is a "struct biosal_assembly_graph_store
> > > (with script "struct thorium_script biosal_assembly_graph_store_script")
> > > and it sends a "struct core_map" to an actor using its name.
> > > The sender does not know the script of the destination. It only knows its name.
> > >
> > > The receiver is a "struct biosal_coverage_distribution" object (with script
> > > "struct thorium_script biosal_coverage_distribution_script").
> > > It receives a "struct core_map" since the sender packed such a core_map anyway. Any other
> > > behavior for this action specifier and for this destination would cause a problem anyway or strange behavior.
> > > There is some sort of contract between the source and the destination for the action ACTION_PUSH_DATA.
> > >
> > > As usual, the source is free to do whatever it wants with the ACTION_PUSH_DATA message since it "controls
> > > its destiny". It could for example do nothing at all with the message and just reply with
> > > ACTION_PUSH_DATA_REPLY. This is how we get robustness.
> > >
> > > > (from assembly_graph_store.c, line 494). It would seemingly need to be unpacked somewhere.
> > >
> > > Yes. I agree.
> > >
> > > The answer is:
> > >
> > > coverage_distribution.c in the list you provided.
> > >
> > >
> > > >
> > > > George
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > George K. Thiruvathukal, PhD
> > > >
> > > > Professor of Computer Science, Loyola University Chicago
> > > >
> > > > Director, Center for Textual Studies and Digital Humanities
> > > > Guest Faculty, Argonne National Laboratory, Math and Computer Science Division
> > > > Editor in Chief, Computing in
> > > >  Science and Engineering (IEEE CS/AIP)
> > > >
> > > > (w)
> > > >
> > >
> > 
> thiruvathukal.com (v) 773.829.4872
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > On Wed, Nov 26, 2014 at 10:23 AM, Boisvert, Sebastien
> > > > <boisvert at anl.gov> wrote:
> > > >
> > > > > From: George K. Thiruvathukal [gkt at cs.luc.edu]
> > > > > Sent: Wednesday, November 26, 2014 10:11 AM
> > > > > To: Boisvert, Sebastien
> > > > > Subject: Re: Causality and ACTION_STOP
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > Hi Seb,
> > > > >
> > > > >
> > > > > See below!
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > On Wed, Nov 26, 2014 at 10:03 AM, Boisvert, Sebastien
> > > > > <boisvert at anl.gov> wrote:
> > > > >
> > > > >
> > > > > > I am looking at spate_metagenome_assembler/spate.c but don't see any usage of core_map. I did see a usage of core_map in argonnite.c, but the map is not actually being sent in a message, AFAICT. :)
> > > > >
> > > > > Spate uses the BioSAL actor library in genomics/
> > > > >
> > > > > Check out genomics/assembly/assembly_graph_store.c line 494
> > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > ​This looks like the magic formula! I didn't look too closely at the inclusion list.
> > > > >
> > > > >
> > > >
> > > > This is basically the same pattern used in thorium_actor_send_vector.
> > > >
> > > > So you can just take that as a starting point for thorium_actor_send_map.
> > > >
> > > > >
> > > > >     new_count = core_map_pack_size(&concrete_self->coverage_distribution);
> > > > >     new_buffer = thorium_actor_allocate(self, new_count);
> > > > >     core_map_pack(&concrete_self->coverage_distribution, new_buffer);
> > > > >
> > > > >     printf("SENDING %s/%d sends map to %d, %d bytes / %d entries\n",
> > > > >
> > > > >                         thorium_actor_script_name(self),
> > > > >                     thorium_actor_name(self),
> > > > >                     customer, new_count,
> > > > >                     (int)core_map_size(&concrete_self->coverage_distribution));
> > > > >
> > > > >     thorium_message_init(&new_message, ACTION_PUSH_DATA, new_count, new_buffer);
> > > > >     thorium_actor_send(self, customer, &new_message);
> > > > >
> > > > >
> > > > >     thorium_message_destroy(&new_message);
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > :-) Thanks!
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > If we do thorium_message_destroy(), does it also free the underlying buffer?
> > > > >
> > > > > ​
> > > > >
> > > >
> > > > No.
> > > >
> > > > thorium_message_destroy() does not free any memory because
> > > > thorium_message_init() does not allocate any memory neither.
> > > >
> > > > I guess this is just because I use to do more C++.
> > > >
> > > > >
> > > > > >
> > > > > >
> > > > > > I'm planning to add these functions to make sending/receiving maps easy.
> > > > >
> > > > > That a good idea. I like it.
> > > > >
> > > > >
> > > > >
> > > > > ​Great! I will give myself this task. I'm going to end up figuring everything out as part of the new version of hello_acq.c.
> > > > >
> > > > > I want maps primarily because I am mapping an actor name to its (boolean) status of having been greeted or not.
> > > >
> > > > This makes sense.
> > > >
> > > > Also, for your information, core_map can never be full although it uses a hash table for the implementation.
> > > >
> > > > Here is the hierarchy:
> > > >
> > > > - core_set is implemented on top of core_map
> > > > - core_map is implemented on top of core_dynamic_hash_table (dynamic hash table uses incremental resizing, so it can never be full and
> > > >                 all calls are still O(1) because of amortization.)
> > > > - core_dynamic_hash_table is implemented on top of core_hash_table (hash_table uses double hashing / open addressing)
> > > > - core_hash_table is implemented on top of core_hash_table_group (right now, there is 1 group per table, but this could be use to
> > > >                 implement a sparse container like in Google Sparse Hash Table based on Donald Knuth's algorithm for that).
> > > >
> > > > >
> > > > > ​
> > > > >
> > > > >
> > > > > > Similar to your view, I think lists and maps are among the most important collections.
> > > > >
> > > > > There is the free_list structure, but it is only used in memory management right now.
> > > > >
> > > > > > I once read a paper that most of the reuse in OOP comes from lists and associative structures.
> > > > >
> > > > > When you say lists, does that include vector. Usually, vector means array and list means
> > > > > linked list.
> > > > >
> > > > >
> > > > >
> > > > > ​Yes, it does in an evolutionary sense. I tend to think of vector and lists as being connected. In modern programming libraries, you often see the interfaces conflated (i.e. lists have array-like interfaces and vice versa). Java Vector is kind of an
>  interesting
> > > > >  case study!
> > > >
> > > > But a Java ArrayList<Integer> has O(1) random access but can trigger a resize of capacity on insertion which is O(n).
> > > >
> > > > A Java LinkedList<Integer> has O(n) random access but has O(1) insertion.
> > > >
> > > > Sure, both implement the List<Integer> interface, but each has a better use case anyway.
> > > >
> > > >
> > > > >
> > > > >
> > > > > ​TTYL,
> > > > > George
> > > > >
> > > > >
> > > > >
> > > >
> > > >
> > > >
> > > >
> > >
> > >
> > >
> > >
> > >
> >
> >
> >
> >
> >
> 
> 
>

More information about the BIOSAL mailing list