VideoPoints

• ••• • •• • ••••• • •
COSC6365 Intro to HPC Fall 2021 - Lecture23_MPI-IV
Help Tour
© Distribution of this video is restricted by its owner
Transcript ×
Auto highlight
Font-size
00:00 Yes. Mhm. Yes. Oh. Mhm. So it is
00:23 oclock, so. Oh. Yeah, I'll talk a little bit
00:38 more about the FBI and then uh yes, the demo uh what time
00:45 left, so yeah, much whenever and it's uh part of it,
00:58 some more sites stuff congress mapping, you have to uh Right, and
01:13 I'll talk a little bit about uh known as hybrid programming, their combined
01:20 empty with M P I. Right, so something just some comments
01:31 some of the NPR implementation as well the open MPI implementation, there is
01:38 room for implementation variations and still needing standards, not all first, not
01:45 features are necessarily implemented by everything um vendor or compiler vendor, so and
01:55 they are features that are not part the standard, but that's okay,
01:59 honest. Whatever minimum sectors a standard is included, distance and this side
02:08 just trying to point out some of aspects, what I think I'm I
02:16 may be important. It's kind of for example, so basically uh the
02:24 corner photograph so it's about the 78 weeks. All right. And
02:33 amount of data traffic between the different , the numbers on the back and
02:43 starting with that it's just the night at seven. Uh for processing note
02:52 which to master processes. The main it's best to go with a human
03:00 01. Yes. Uh First uh and the next two lives that you
03:12 to the next hole just taking this it is uh hurried up into camps
03:22 uh you get that order so I to do that. Then the edges
03:32 in this example and number honest. there's a lot of messages uh so
03:40 we take the different levels efficient, possible the mistake. No terra I
03:48 too that things are so The process on one so zero and what so
04:07 find a way but so There are or 55 years old. They were
04:18 70 91 and 1-2 is uh 7 40 ft. So uh in
04:31 okay. Well don't oh onesie, sorry. Long too 46, 6
04:42 take them off what? Uh to this time to stand up the edges
04:51 disaster than just. Uh huh. nos another one of these things.
05:03 six. Well, you know, 44. Didn't learn about that,
05:16 what the market traffic. Uh And some of them then there is
05:24 I support that ways of doing the sentence even progress those in order to
05:33 numbering and numbers is trying to do for a number. So family optimized
05:44 . So you can obviously see that numbers going on that most of the
05:53 , references or data at least, processes are local memory references and things
06:01 on point. It is zero five is 71 - one. Uh
06:15 right. Are six Not for So that's like Uh huh. We
06:31 covered. Thanks. Still in This is what you get is now
06:38 feel safe for me. Yeah. be it Don't be scared significant reduction
06:46 the load onto the communication network. that's part of the problem or challenge
06:53 terms of no mapping is to since an implication that protects the people with
07:01 part. So to get a good of your code, you want to
07:07 a good market. Um And so name of uh system operates.
07:23 System uh if you use a measure the craft and Amanda describing process
07:34 then the system they may have a way easier way of thinking of the
07:42 processes to note I guess in this is another trigger one. But it's
07:48 same example example. Yes, that's doesn't mean that the water very
08:05 Okay, so then the graph describes biological on the there are references between
08:13 . So, in the name of uh traditional waters and mr riddick.
08:22 . Uh huh. Sorry. School . So yeah. Mm hmm.
08:29 bastard. The first of the year I was talking to That's good.
08:42 So funding uh right, supports the winter cold. And our processes please
08:55 up from each other. Mhm. know, they're just feeling kind of
09:03 . So Okay, so this is application. That is correct. Process
09:18 . Now this interactive process street. when it comes to finish, that's
09:35 in the name of So here is , that's the they got the representational
09:42 had from both. Well, two on the lecture. So the medication
09:50 for I mean these are you they got traffic between these two
09:56 Uh, anything about it doesn't look that and your master onto note that
10:10 way. So the process terrible across fun of the transformer compositional And then
10:19 four and 5. Get smart awkward far through three. So this is
10:27 and that's lightning. Hello the notes the machine and the custard. And
10:37 usually don't have any control over So that's the nose gets address and
10:43 so happened. So it's usually start zero up to whatever number it takes
10:49 cover all the notes. But you bet on an idea to get an
10:55 . But the street let me know there's a sign and I when you
11:02 the system so this, this is device uh, arrangements get back on
11:09 investigation. Uh, the pastor It depends on oh, what the
11:18 of different data references that have Where are they? Nothing is what
11:27 are. Okay. All the processes the notes and the machine and all
11:34 difference. The president minimize the overall for all the basic broken up trafficking
11:46 . That's one way of trying to enough of that. But you know
11:51 about the connectivity between these four walls mean to something that actually is because
11:59 you also once you know the fact the nose and then directly after that
12:07 can also signal contentious in the next this point there is no notion what
12:12 network is and what the confession. huh. So in order to do
12:20 think you need to know? So know they want the catholic state of
12:27 study involved and they all activities in to actually buy something attention. That's
12:35 issue. Or if it's uh regarding network our capability to make not necessarily
12:43 that's me just walk to the university we lost the track. Show me
12:50 became they need to process in fact grown up, correct. So that
12:59 that differences are demanding all the the non blocking. Yes. Oh
13:09 . So if the diarrhea historians point but it also done depends on how
13:17 the expertise the principal, But it's in constant. So it depends on
13:26 six the internet you have to uh know, you're very tired.
13:37 That's quick. Mhm. So so here's just an illustration tribute
13:55 Uh I think properties we get along other greens, both clean and prepare
14:08 about 13 certain market potential. Whether know, in this case there are
14:18 things that mhm 10 points in the of the Children. Oh thank
14:27 That is something to the complexity, on the running out by itself that's
14:37 take care of the establishment. FBI generally does not. But the
14:45 person so. Right, wow. uh so the switches uh speak softer
15:00 for some network censors and his Okay, look at the right pattern
15:06 they know about that's multiple. Important. They're trying to minimize context
15:13 it's usually an opportunity that my implementation machine that died the NBA has the
15:21 to the source and destination and then it over a bit mr simmons into
15:29 on something and there are others internet , you know how the messages come
15:35 point A to for people. Uh . The other source this and potentially
15:45 priority what kind of features but it know about the past few days in
15:52 and how to do that. That's network itself. Yeah. So then
16:08 Uh huh additions about cigar stem. telling her it's also something that's
16:19 It's coming in terms of how to the mapping. But as you will
16:25 on the side matter of fact. huh. The math intensity quite
16:34 So there are lots of features of where things ends up because you
16:40 as I mentioned moving there is the critical part of performance uh you know
16:49 started in the urban actors. Of the stress of memory bandits and to
16:54 a problem. Many applications and then a final entity use and there's another
17:01 that is weaker than memory busses, is the Pizza Express and that enables
17:08 to then there's another piece that is a weaker and the peace that expressed
17:13 that an expert itself and that's why of this mapping is important and that's
17:21 there's so many features provided. So I don't think this, you
17:30 want to comment on some of the that have forget to mention that.
17:36 uh and and those expenses, remember was saying uh because it's kind of
17:46 all the options. So that's you know, where else to and
17:51 is opening Yeah. Standards and most supported the standards or even if it's
17:58 internet, you are somebody else's uh a good starting point to save both
18:05 the standard and they're not gonna find tweets, various companies have done but
18:10 particular implementation um mm one having lots things. So there is that's
18:20 I'll come back to this notion of in the next five to but this
18:25 even more yes, move there is mapping, I don't know that of
18:33 processes elevated and with examples general and it's not affecting the binding. Stop
18:48 the difference of and all that not about it. It's not the
18:57 So it's different two nights ago. there's uh as far as pretty
19:07 I would say uh of what the of these things are and I think
19:12 through examples, house gets a little better understanding of, of course I'm
19:18 a different, well ways of controlling things sends up. Um and I
19:28 from that, think about it because the NP Iran refused to, the
19:36 was sort of an FBI program, then we also had this notion of
19:43 went to ask for money knows that controls I'm ahead processes. We want
19:50 know what it's saying. So generally run depends on a scheduler or resource
19:59 , life store. So it's and . So in terms of getting information
20:05 the cluster and the notes and all that. Yeah. So the,
20:10 on what they have told someone that want and uh Cool. Yes,
20:20 . Uh well a little bit about , but then absolutely not, but
20:33 person and that was fine and the , but this that's immediately what
20:40 Uh miss scenarios where what the defaults all fine. So I think I
20:51 some sites mhm later on. so one is specified holes and there
20:59 many different ways to specify holes. you go. I'm directing or but
21:04 least we cannot define, I'm not it's not serious. Uh the people
21:14 tactics and is that french no Venezuela for what happens happens. Um and
21:34 principle the standard last year or not a long, I would say all
21:42 the different caches um, and they all day support that level. But
21:50 things from our reference in court uh . Uh, some of them also
21:58 cool. The one that is but uh, I mean in some
22:12 that's got to be back four harbor the best stuff. The reputation,
22:22 stopped me and different than uh, B I know and the administration's there
22:37 the default. You don't name anything start trying to figure out what
22:43 It's what is a good thing to the target on that machine. Whether
22:51 a court makes sense over the subject sense of kind of you don't need
22:58 . But the good meaning or so conduct. So no. And I
23:14 want to miss something by default. are okay. And the icon allows
23:26 feel more process is that there are that cores or sockets and so some
23:34 them, whatever the entity is, be oversubscribed. But if you don't
23:39 that, wow. Yeah. There's oversubscription then so much that I'm sure
23:50 they're not type of targets to use them acting there is plentiful enough.
23:55 there's no, it can be higher the number of processes but we cannot
24:00 more processes than targets to know all subscriptions. The other one that is
24:08 to what the last thing was that was all right. Like this.
24:18 vocabulary wants to contact spread that way will pack things or stupid things as
24:26 as possible. So bad answer fred Otherwise, if you don't specify diverse
24:37 version other than magic. Uh The number of processes to wow The
24:53 Yeah. Mhm comments is in My comments I think uh over should
25:05 mount everyone that for over subscription. generally it's useful when you have processes
25:14 may stall for some reason and in time the other processes can do their
25:20 on the functional units. So let's if you have a process that's waiting
25:25 I. O. Or a file or whatever internet request or something in
25:30 is you may want to have over so that some other process can utilize
25:34 functional units to get better efficiency. that's none of the use cases you
25:38 have or subscription both arms. It on your application, how you have
25:46 it. I'm other way of thinking that if you if you have multiple
25:54 units and each process does not have work to utilize the all the entire
26:00 units, then in that case you want to have over subscription as
26:03 That's similar to the reason why sometimes have hyper threading enabled on some of
26:08 systems. So that would be one reason to do that. So it's
26:14 necessarily that every time you will benefit over subscribing but depends on your
26:20 Yeah. All right. Always Um We're telling him to walk on
26:33 . So again well, well, subscription is an option available to you
26:44 use that or you can disable So if we is everything that is
26:51 if you just uh straight up for and then most so then maybe I
26:56 give you another and we I will not run your program. It will
27:00 it but it will not run your that will give you more processes.
27:06 whatever the target is not to force topic knows. Oh, so if
27:13 deficient and the resources come back to processes over subscription. Yeah. Because
27:30 at compile time the compiler doesn't know many cores and no treatment. That
27:38 what is finally. But then the chapter resulted from only when you do
27:49 , you specify how many processes you . Mm So there's younger sister kim
28:04 that. Go off man. No is only when you ask for resources
28:11 than what you have available. That's I have my resources. But uh
28:20 need uh Yeah. Then you need cancel your job and then request,
28:34 more resources. And then yeah, a grand time. You cannot uh
28:41 the number of processes once you have some number of processes in M.
28:45 . I. During runtime you cannot or increase their processes. Some processes
28:51 end up having no work and they sit idle but they will still be
28:56 in the runtime environment, they will get deleted or whatever. So uh
29:08 . Yeah. So this is sort this is so fun. Well defined
29:18 . Oh, also uh No. uh to they're supposed to one each
29:26 of the books. Uh huh. number five. Not my process,
29:39 know. Uh whatever I think some about on Yeah. Oh yeah,
30:00 fun. Lovin breakfast. Oh, are process experiment offered. Uh We
30:15 to know 72 sockets indefensible and 30,000 targets for so that to have the
30:28 instead of depending on the person. , no. People talk about four
30:36 them. Thanks. And the other told you look. Uh huh.
30:50 me, you know the earth for . Okay. The contract allocation
31:03 Oh, And the next 4? , the only six years after.
31:12 . And I'll see you get So it doesn't try to those
31:17 But there's something that because Back in 60s, both B- two C um
31:31 here is kind of the example, one word for green. Right?
31:42 there are always subject to pretend. the first thing he does in a
31:48 From four of the time for the responsible for processing the first book holidays
31:56 order agency. Uh and then I'll more processes and always decide market.
32:05 Yes. You uh almost from the here, that's why you may get
32:16 heritage that still to this day. is the work over. No more
32:27 4 41. Okay. But with this home because that was located
32:37 the suspect. Cc Yeah. So the next thing was sort of
32:49 and blind attributes and one comments this . So this central finance actually job
33:03 , the Hamburg scientist and he budget interest that are particularly good for their
33:12 architectures, but many of them. the first few sides here for instance
33:21 the opposite. So let's see what sites here. So it was met
33:26 court to the extent that so in case and so sick offices for so
33:39 words no meat. Uh those by core uh, take them in
33:47 to transport their ranks. Don't say moves up beef. I don't know
33:56 socket. Uh Oh, it's all course. Uh First soccer.
34:05 reserved for the first process. yes, ma'am with that. And
34:13 the next one that report next Okay. Buy the product. This
34:20 from dropping version of the school. he's uh do something before processes all
34:28 . That will be just a quick the dots in that diagram is there
34:38 got send up one above. So means that there is hyper threading enabled
34:43 each core has tried to hardware trains . So that's why there's two easy
34:48 you don't have hyper threading, you see just one B or whatever symbol
34:53 implementation uses. So yes the the corps. Yeah you can If you
35:03 map by hardware thread then you'll see one of the bees and write one
35:08 . I don't know if you have example. That's what happened. So
35:12 , the other one. Oh come 55 right now And still not but
35:30 right. four. That's quite the . Um Okay. Different order.
35:38 smooth. All right. The first one of them, the second
35:49 It's very tough. Find the ground quarter This is And the findings from
36:02 place. Water four. So that's finding explain the process. They're always
36:16 the process between the chords in Osaka succeeded. I want to find it
36:23 more. So then. Oh even . Uh huh. What? I'm
36:39 ? Mhm. Uh The most fire . So and that is uh words
36:51 first. Okay. And then vaccine for the expert has been all levels
36:59 suppose the top position. So it's chose you know all kinds of ways
37:08 trying to Allocate things and depending upon application one or another maybe better.
37:15 I think they have some others. like discussing that a little bit uh
37:23 on this one and uh well, for four. Uh huh.
37:42 And Oh this is B. Found the the other one. But it's
37:54 many of the resources used for across person? This is Excuse report.
38:10 Fun. Oh awesome. Based on . So typical for what?
38:27 mhm. Who is this? Uh prophecies. I'm so hopeful. Stop
38:40 . Uh huh. It's all kinds options and and I think business
38:48 I wanted to common time. So is the way to think about began
38:56 the architecture it looks like So remember they on some processor from protectors?
39:08 is so thin stock it. So started the uniform member channels. I'm
39:18 if you want the tax might stand and you want to to lose both
39:25 because you make more money. And the number of members that compared
39:33 using one socket. Yeah, three sockets so try to But it's
39:43 seven and it's optimized. Mm. anyone work here? When you say
39:56 guns? Yes. Okay. 2000 approach that somehow we mentioned,
40:08 I'm gonna usually mentioned that the domestic in an implement reverse tickle too when
40:18 was that similar to those on the sir. Yeah. Okay. Uh
40:26 agreement. Well, you're in the 1. You're mapping by socket but
40:33 the bottom when you're mapping by right? And I think the comment
40:40 , you don't want your alternating processes go to the next socket but rather
40:45 in the same socket. So in case you can use the slot as
40:50 mapping. Uh We don't switch. is large. If you have only
41:05 field we say different. That stuff the meaning of stop it's something doctor
41:17 an FBI content system son it could our reference or you can stop
41:27 So all the thing this case the before by what the end result is
41:41 you start to learn the meaning of . So many things that happened.
41:48 is uh it's not so that there huge of loss believes process. Yeah
42:04 of course of course this the So I mean but when he was
42:17 in the demo we got the same in the same fashion. I don't
42:22 know what that discussion that's so what trying to say is that the definition
42:31 slaughter is implementation dependent. So in case slot meant approved. So that's
42:38 it's looking similar to what you get doing map I corps. Right.
42:46 that that actually were the right that it's an imaginary concept so it can
42:51 anything either it can mean hardware threat socket or anything that depends on the
43:00 . So you can see in this for the world but everything has a
43:08 of the one the one suffering acute you kind of only about half of
43:15 family. Uh people want to work the next slide but this concept we
43:21 on the other hand sometimes these things anyone they they optimized because not everything
43:32 on the same subject. So Have good day that we use the cash
43:37 up 90 spouses. So it depends the application, whether you kind of
43:44 to stick things from every standard is a mystery but they may want
43:50 nothing. But it is that the is that we want to. Mhm
43:58 huh. So that's why I have those knobs to manipulate because it starts
44:03 clear that uh from whatever they run knows about what's happening and they always
44:11 they can figure it out. Oh , so for example and see what
44:22 not over here. Oh yeah. , oh I see that. So
44:34 has additional options but before mother component on the defense, how they built
44:42 own processors. That's kind of very And the modular even in this blue
44:49 itself. Yeah, so I'm not to dive into too much on
44:56 You see something you want to So I just have these examples for
45:02 . They're interested to take a look and see all kinds of options that
45:07 available for us specifying basically what jay gets allocated to each process and where
45:16 end up and if you find them I want to, you know,
45:22 open MPI I had a bunch of that shows that you can get in
45:27 case either binding the core or hardened the socket so that you know,
45:32 different options for the always simple things that could be good. Uh If
45:38 not the only one running on the because Mr may allocate other jobs to
45:44 of the courts of europe. so that they can dependent and it
45:49 move it around except the human benefits could also be affectionate. So mm
45:59 there's, you know, different happens uh yes. Okay, so next
46:05 I would do so anything more about ? Nothing was done. Something
46:15 Just a great general comment. It's what they've been doing in their assignments
46:21 uh open MPI or single credit It was necessary to understand the memory
46:27 pattern. That's kind of the similar that we're trying to achieve an open
46:31 as well, but where you send , where you allocate your processes has
46:38 significant effect on how much data communication between the sockets or across the
46:44 So whenever you're trying to do you take your project. So it's important
46:47 understand how the memory is accessed in the processes and what the data communication
46:54 is now and that would really help to get good performance if you can
46:58 and map your processes accordingly. Like of the not mapping examples that was
47:03 initially. Right, So next a comments about combination of np I and
47:17 mps ascending principle. Iran you have FBI processes running on the same,
47:25 . But then process the process communication much more weight than with the friends
47:32 opening piece. That's why it I've been programming, developing story the
47:42 that's what's happened. Data between processes different process source or else. And
47:50 within the nose to the memory stand open every day for that. I
47:57 have any but it's but they both once again the process is I know
48:03 we just talked about but some of point of view is standing right
48:08 right 61 I really like to know then there's empty. I there is
48:16 easy that's the notion program him on decided really want to silence place,
48:28 know, secrets. Yeah. Uh then there's a few different ways on
48:39 . NPR those we are used to . I can open and eat um
48:50 available. So here is against one the pictures. Again, the combination
49:00 this process that can be friends. uh no. Uh the Beatles you're
49:12 between. No, then there's all of this uh victory. The extent
49:22 which Japan I told Uh one French allowed to sleep. FBI Cool.
49:38 one safer. Mhm. Oh, the main threat to have an
49:49 So that's a friendship and it's the one that please for a second time
50:03 uh please for all the possible but the giants for different the city itself
50:14 with it. Uh And that's why have these options. So here is
50:19 a single, That means just one 1st. No. Mhm. one
50:26 at the time or you can do bundle in this case the master for
50:31 exclusive and then they can spawn a of threats um and come back transporting
50:40 again. And this is the cereal based upon at the time version in
50:46 of the FBI calls and it's just I said everyone can do something but
50:53 more of a challenge to keep track so it affects which one you want
50:59 use, depends on your application and on. But it just shows the
51:05 . Uh huh options that from us of course developed tickets. It's more
51:10 a challenge both on the programmer to sure that you can always get the
51:15 answer regardless of what you do as as for and the irony of system
51:23 make sure oh the rules of the are kind of Yeah, questions and
51:35 comments and this is just pictures in coma and we'll soon stop online.
51:44 of them are some versions of Uh How did it? I'm not
51:50 we have the every emotion. So is just and this I think this
52:03 and damp black is on that Yeah, it's so you if you
52:08 to play with it on your Um And this is another version of
52:16 trade and whatever and then because you also pointing out that when different groups
52:22 issue things, things are very critical mature. You get the correct
52:25 So because mp doesn't necessarily which of in this space um broadcasting on the
52:37 to end the protest operation. So on um you don't want the threats
52:46 be confused. Who's doing what or . P. I. To be
52:50 . That's still what you want to . Some of the exposed an incorrect
52:54 and take a look at. Then also they finally incorrect example of trying
53:01 see her in order to And this definitely NPR so let's freedom for multiple
53:12 , history material pulse. So this the government, the demo.
53:25 So national too. And that's I a few more slides here. But
53:35 just so you take a look at think there's something specifically has chosen this
53:42 bit performance. All data but now professors system data structure. Okay,
53:56 me just take over from now Yeah. Sorry. I didn't prepare
54:01 the uh everyone's but the cool examples like okay, if you guys can
54:08 with. Yeah. Yeah. You awesome. Uh huh. Yeah.
54:33 that size good enough? Yeah. just show so these all these examples
54:41 you see here in this directory they're on, I think under the content
54:46 on blackboard uh holder for your That's why it doesn't make fun.
54:53 just to download them to make Uh huh. All the developments not
55:02 mm hmm. I'm sorry, but very simple. Uh huh program.
55:13 this lecture. So basically NPR program the five points there. So uh
55:21 uh decompose your entire grid into uh three smaller grids and distributed across the
55:29 . So, uh in terms of , you also have a duty grid
55:32 processes. And each process in that gets one block of the matrix that
55:38 supposed to work on and stand. as it's called the nearest labor communication
55:44 . So when you decompose it, all the processes needs to communicate with
55:50 neighbors which are as according to what call north, south east and west
55:56 for the for the reasons that the processes need to perform their own
56:04 So initially nothing special here, it's mostly all the FBI initialization calls doing
56:11 parameter parsing here. The main thing from here here you letter mind what
56:20 you're for neighbors are based on your drank. And so your P.
56:27 . Is the number of absolutely There is a P. Y.
56:32 well somewhere that's the number of facts the Y direction. This is
56:37 Uh Good sizes were always a simple list oppression that allows you to compute
56:44 if you have a neighbor and north or east or west if you have
56:49 or not. And then based on you communicate with that particular neighbor that
56:54 exist. Uh these are your 2 Great old and old and near 11
57:02 your computer results, that's the old then later on used to have uh
57:07 progress for the next generation to We and like 1000 iterations which will
57:11 on one sec. Oh yeah the there as well. Mhm.
57:19 So in this in this example the it's a very simple one. You
57:24 define uh these offers which are contiguous for each side. So as uh
57:33 means for each of the samples, back your data for each direction into
57:39 particular offers and then send it to neighbor that actually requires it and I
57:45 our prefix uh the receive buffer and that's about where your neighbor sends you
57:52 data that you require for your own . So in this case those are
57:58 minded your stuffers, we're not using data from the from the internet.
58:04 then simply here you start your inflation the jacoby uh you act your data
58:12 each direction into into the land And then uh what is here that
58:18 use these non blocking center received for ice. And and I received.
58:23 first you do I send for each and then you also also, I
58:29 the directions we wait for you to the data from your neighbors as
58:33 And you call the FBI at all is basically a synchronization point to make
58:37 that you got the data because you know for these non blocking girls if
58:41 got the data or not until the point. And once you get your
58:47 are you perform the updates on your . That's very simple implementation for this
58:53 be uh the communication and they need do your computations. So I'll just
59:02 go through all the old examples and I have some numbers to show you
59:06 you the uh after that. So than using the contiguous buffers, you
59:17 you can also use these derived data from NPR. So in this case
59:22 you remember from the last lecture we uh they don't like called FBI convenience
59:27 allows you to define data like that a contiguous buffer in this case
59:32 X. Is that the number of that we want to send? What
59:38 and the I double is your data . And then you just give your
59:41 type and name in this case. north south dive which allows you to
59:46 continue uh communications for normal for eastern side did I strided if you
59:54 So there's because we work with row format and sees or police investigator destroyed
60:00 . In that case we use these protector the data. Again you have
60:07 size uh number of elements that's one unlock and then the strike between each
60:14 and the glass on us again the of your of your of your local
60:21 . And then I I actually actually the uh in your bed and give
60:29 name for that data. And then this example is pretty much the
60:33 Uh you need to make sure that commit that you're right, you can't
60:39 it and that's pretty much the So now rather than doing the packing
60:45 unpacking of data and those continuous government the previous example you can simply do
60:52 I send with your delight data rights than because they're doing that back in
60:57 biting into the country's coffers. And you perform your computations on the on
61:03 data that you have. So again using the non blocking centers us the
61:10 and and I received here man does uh can anyone think of one simple
61:23 that you can perform? Uh For this expensive example we're doing
61:34 We do and then we do the on all grids. Simple, very
61:39 optimization that you can think of. God. Well the simplest example uh
61:58 optimization you can do is imagine that these moments. Why are you performing
62:04 uh these communication operations? The only is because for the alleged elements you
62:10 data from your neighbor because uh that has the The elements that you require
62:16 do the five. for the inner of the grid. You really don't
62:21 a needle in your neighborhood. So this case you can simply post your
62:26 operations using the icing. And I while the communication is happening you can
62:32 your computations on the inner part of world so they have eaten. You
62:36 need to wait until you get the from the outer part of the tribe
62:41 where you can overlook the combination of . And once you have done your
62:46 computations, you can check whether you got the data by having uh about
62:52 in the end and then perform your from the air on the outer part
62:56 the. That's supposed to give you speed up but that also depends on
63:01 your network speed is and how you map your processes. So you can
63:05 get some speed up. You may get it application as well as implementation
63:12 . But in general I guess you probably get some speed up the
63:16 I I did not get a speed . That's why I'm making a
63:22 Um Yes. So even in this we were using the that they arrived
63:28 . Right? So again the right to live as I've noticed from my
63:34 , it's mainly a way to allow to have easily program easy program ability
63:39 your program because when you have only numbers you need to figure out the
63:46 and then you have to figure out indexes of elements to copy it,
63:49 data to. First to the buffers then to the senate you have to
63:52 through all that assault the right data you only need to do that once
63:57 then you have the right data type then you can basically just just refer
64:01 that there are data types can be to handle all that conversation for
64:06 So again with my measurements, I not see much speed up using the
64:11 data but I feel like it's more a program ability uh future for the
64:24 to come from the next project. is this I mean so the internet
64:30 uh going up quickly so you wouldn't to do. It has more number
64:34 things but double overlooking only become time most of the, depends on how
64:46 you have distributed your, how much your ancestor is. If you have
64:55 a balance between the work for process the communication time That happens between the
65:02 processes then you may probably see spr overlapping communication and computation. But if
65:09 is the bottom line and whatever you in terms of computation, that time
65:14 still going to be a part of total total execution time communication.
65:21 But how was dinner time? uh in in the one time it
65:29 is doing the same thing that people manually in the one time there should
65:33 a function that's actually going and getting the values whether instructors tried and Australia
65:41 and the back and that's probably what's on. But it's likely to have
65:45 it and some more optimized fashion than it entirely like you would with a
65:51 problem. A more optimized. So , I mean comments from you.
66:07 . And they didn't. Right. huh. I would think in
66:18 Okay. You know what time it . First of all the ladies,
66:28 one night. Mm hmm. And the bible is not on the let's
66:36 . How has on the genes that up copy is picking up?
66:50 I agree with you. Yeah. the 19 it's never enough. Uh
67:02 . Just do one thing to remember to change it somewhere. Okay.
67:15 I hope to perform different. Here's problem with the First nations, things
67:22 that. Yeah. On the news the giants. Yes. Yeah,
67:32 the one that I showed you initially first example that I shall initially.
67:43 . Oh no, that that uh you do a scent for each,
67:50 other. Yeah. You don't generally . No, I did not do
68:03 problem. Yes. So rather than the data first to a buffer and
68:09 standing it, we do a stand each of the elements that you want
68:13 send in that case the same operation an overhead of doing it.
68:17 So it needs to talk to the I deliver whatever networking there there is
68:22 then send that one element. So case of copying the data into 20
68:27 before and then sending it in that you only pay the overhead fee of
68:32 a sense only once for that environment you're doing one element at a time
68:38 sending it one x 1, you that overhead of doing a scent once
68:44 each element. So that overhead adds the total execution time and then you
68:49 see a heart condition. Well seriously? Yeah. And then the
69:01 example I have heard was showing you the remote memory access is that's the
69:10 , that's the one side of communication . So rather than doing uh send
69:17 receive, you know what and um , so where is are you
69:27 Yeah, so this is basically from lecture you have to set up a
69:31 uh an FBI for doing this uh side of communication and then you can
69:38 put so in this case you don't need it centralization point. But since
69:45 application depends on getting the data and on it, you need to have
69:49 organization fund in this case. So this case also I did not see
69:54 uh speed up, I'll show your in a moment this other thing in
70:01 of yeah been using these one sided , remember you need to define these
70:08 which are remotely accessible members faces on process so that you can do this
70:13 won by determination because one side of remember the receiving process does not post
70:20 receive operation. Right? The center simply puts the data into the into
70:25 remote, remotely accessible memory of the . So in this case each process
70:30 need to create a window there that can assist. Right? And there
70:40 some of the numbers that are and got pretty much last night. Uh
70:46 these are on the on the bridges note I got access to two
70:51 there's 1 28 processes on each showed that's equal to the number of course
70:57 their nose as well when uh Uh I would go to 256 uh piece
71:06 these numbers that you see are for by poll. So each the next
71:12 goes to uh to the next physical in the in the system. And
71:17 it's also blinded only a thinking thing we cannot move between any other
71:23 So one very simple. Mm mm hmm. one Very simple observation
71:32 can make here. Is that The scalability of your this around
71:39 Yeah, as soon as you uh one Grand EA processes, we are
71:46 only on a single line here so does see how good balance between the
71:52 or process here. So the 128 you pretty much see uh the execution
72:00 to decrease for the entire execution, as soon as you move uh to
72:08 process easier that you're using our tools then there's a communication link between those
72:14 months. Then the scalability starts to their Because I have introduced one model
72:21 your own time at center. what else? Yeah, that's
72:34 Right? Yeah. And yeah the thing that I want to mention
72:43 is the map the importance of So now on this, right,
72:49 timetable I'm mapping by note, that each consecutive process goes to the next
72:58 uh in your other patients. So got, you've got who knows The
73:01 zero stays on North zero Process one to second note as the long
73:07 Right? So now, because this the nearest nearest neighbor communication application now
73:13 process zero has to go across the regard its data or send send the
73:18 . Now I have to travel a distance to the data communication and I
73:23 get all the results before getting to in here. Now you can see
73:28 the scalability now drastically uh increases as you move towards more processes. And
73:35 if you compare here, let's say processes here, which is the first
73:42 here and then the full processes with of mapping, of course You can
73:48 see there's almost four x decrease in atmosphere. That's just one example to
73:54 you the performance or Yes, these execution banks so Yeah. Large.
74:05 . Mhm. Talk to you. . Exactly. four point they
74:14 Exactly. Yeah, that's mhm. showing you the importance of mapping.
74:24 has to be careful very well, much even mapping in open MPI or
74:29 uh Both only that some right now everyone. Uh huh. All this
74:44 know, although spoken on network. right, well that's fine,
74:52 Data references of memories prominent. So the point seen in this case.
75:04 . Oh obviously everyone else was No, okay. On the right
75:15 that time. Rare point of Oh right six. The association and
75:31 that's why I'm uh both models are used uh basically involvement, processes,
75:41 on the outside. On the right is all the time. He this
75:48 tough. Uh huh. Yeah. . But still it will be larger
76:00 that. I don't have the Yeah. Yeah. So called liberal
76:07 . Oh yeah. Okay, window just a logical term for for saying
76:34 referencing a remote memory location information. just a logical grounds. So that
76:39 of the patients would be on another , it would be on the same
76:42 . It would be in another The window is just one.
76:48 it could be on the same notice it could be on the same socket
76:53 in that case it really doesn't make sense. Yeah. Yeah, but
77:02 that's the flexibility that allows you, can you can still do it if
77:06 application is as such, that it make sense. I'm not saying
77:10 I will never make sense. I'm saying that it depends on your
77:13 Whether it makes sense to do it not. Window, Window, window
77:21 be a window is a remotely accessible location. That's right. And those
77:30 uh Windows are known to every process oil in your runtime environment. That
77:35 that inside any inside your environment can single socket. A one note multiple
77:47 can be anything. So, all , they know about that remote memory
77:54 . That is blinded to one particular . So. All right. And
78:00 you can the other processes can come basically say that I want to put
78:03 daylight in your remote memory location that have presented to the end. Because
78:13 a there's a concept of having remote and having a section which is
78:18 private private memory. Private memory can be used by the process that the
78:24 memory location is the Yeah, the location that that process exposes to other
78:29 so that they can communicate with that . Oh, alright, sorry,
78:35 just have two minutes because this will Yeah, because this will be
78:42 What? Yeah, but I just to show one thing that's something that's
78:51 you're thinking. So because uh some you might use Mp, I just
78:58 to show how you can use the . But uh now with the FBI
79:03 that will be useful for you. again, if you the uh loading
79:07 that module. Mm hmm. Then can have you need you need to
79:16 your exile and in this case that be the main file that has the
79:24 the FBI option uh because that's that's one that's been set up. So
79:31 have this Mhm. Uh and then these are super sports programs that I
79:38 here for sensitive, so I'll use uh how C X C X
79:47 Which is the uh C X x plus plus from my other. Uh
79:57 . Yes. See, excites And basically that's the same command that
80:03 there for any of the programs here then you do uh expensive FBI dot
80:11 . And then whatever your output is beautiful name, whatever you want to
80:15 it, I want to be there now. And then while running your
80:20 , what you need to do is have to give all your mp irony
80:25 parameters for, So let's say you to support system for your program That
80:30 specified 1st. And then it's similar what you did for uh for your
80:37 programs. So now, exactly. then whatever thou uh options you want
80:42 give it here uh since you mentioned run, you don't need to set
80:47 p flag to say to specify to . But there's a way to do
80:53 as well. And then it's complete your program there. The the only
80:58 you'll see when dull generates these metrics you. So, again, if
81:04 remember if you have multiple metrics separates , one directory for each metric.
81:09 , if I Surely one of Yeah. So I ran this uh
81:20 example, four processes. And until until open MPI you've been running only
81:26 process. You've been doing multi credit , but not multi process program.
81:31 multi process program drumming. So that's after now you've been only seeing one
81:36 that was generated by the data for the threats that were involved in your
81:41 In this case, because I have processes here. You it generates four
81:46 files. Which process. All But reading it. Uh still the
81:51 . You just read the graph. that opens the uh all the profiles
81:57 each other in the process. I think there is laid up for
82:00 the processes here. So if you from here is the Not 0 -
82:07 . Um Go down. No, 11. basically culture. The process
82:13 not. So, I don't get in that and I'm not too.
82:19 then in the end there is no . We had those four processes which
82:25 responding is No. And then towards end you have the summary or uh
82:31 the uh all the processes. So get the total for somebody for each
82:37 for each function that was involved. then you also get the mean
82:42 Probably a little process is one of metric that you're running right. It
82:48 you details about all the FBI functions would bother if you have any other
82:52 are doing in there just like they're of energy. So I just wanted
82:57 show you that you can receive I know sorry. Now if you have
83:02 questions go ahead, I just want finish it up before me before hybrid
83:12 . Yeah steve right. Yeah. that case you need to do dash
83:17 uh okay, A and B. also you need to do the FBI
83:22 first The 500 won. And then guess each of the four profiles that
83:28 showed you each profile rather than having one section, it will have multiple
83:32 for the uh Okay, that's pretty implemented. Thanks to you. It's
83:51 lot, you can be sure that kind of thinking through. Sorry.
83:57 happens when we think about this. well if you do it then it
84:04 you uh all the other processes, how much work? It depends how
84:17 No, no, no, it on how much how much work each
84:22 . So if you have four processes like I have uh process one Process
84:26 , Then, let's say 1000 Process students, some other number of
84:32 , and then the response will be different. Right? I'm not going
84:36 get the same every time unless each did the required amount before I'm
-
+