© Distribution of this video is restricted by its owner
Transcript ×
Auto highlight
Font-size
00:00 hurtful. So, uh, so . So it s O and I'll

00:11 a little bit more about the composition data structures or data sets. You

00:16 , for clusters or MPP, where a communication network going to connection network

00:25 you need to decide what goes where for anything that requires cluster. The

00:34 talked about on Monday as well as is, should be valuable to

00:41 So that's why I kind of, , include them and the classes.

00:48 sometimes people, um that a doctor other disciplines, they don't quite understand

00:55 their tools for petitioning data structures. even though we don't have an assignment

01:01 you to do it, But so is for your general education, Let's

01:07 so. I was about to start Specter by section, and then I'll

01:12 about that on the other bullets on slide. All right, so this

01:18 again The purpose is to carve up data using this idea of an owner

01:24 ALS and most of the computations air whether data is But since you're solving

01:30 problem that doesn't fifth and an typically or for other reasons, they

01:36 more complete powers. He has spread out over a number off notes that

01:43 them recently load balance. And I to, uh, minimize or of

01:50 excessive communication, because the communication net , even slower than memory access,

01:59 by 2 to 3 orders of So that's the whole of the

02:05 All right, You know, respectable section I showed this slide and

02:09 about the string plucking a string in various again loads are illustrated here

02:15 um well, the vibrational states, you like or again notice the

02:19 I used him alternatively, and talking these things and some of you that

02:25 familiar with argon analysis. For those you, hey, may make

02:31 The point is that the middle one and I says lambda, too.

02:35 string pictures is the one that has of been used for partitioning graphs in

02:42 . And, um so here is a little bit that that did not

02:47 talk much about, But if you a graph fun conform. This called

02:52 matrix that may be familiar from some topic about electrical engineering or physics or

02:59 one also uses discreet representation. Sometimes . I want to try to

03:07 and I have a picture on the slide. But basically where the

03:11 my tricks e want to set it . So one has one role for

03:15 note in the graph. And then has one column for every,

03:19 edge in the graph? And then because you get to interests in every

03:27 that corresponds to non zero interests. should say that corresponds to the notes

03:37 the end, ends the edge. then there's another one that,

03:43 lactation matrix that is used in That's it for doing draft the composition

03:51 it's a symmetric methods and metrics. this case that one has. There's

03:56 same number. Off rose and colors house show the next kind of slide

04:04 they illustrate these two things, so too trivial graphs kind off on the

04:10 hand side. And then there's the . Corresponding incidents matrix in the middle

04:13 the slide, and then we have loved policy and on the right hand

04:19 . So if you look at the line graph again, there is a

04:25 notes. There's five rolls on, , therefore, edges therefore columns.

04:30 if you look at the recall, then there are two interest in each

04:33 of them that corresponds to the end off the corresponding edge. So first

04:42 number one there is a blue one it's right on track. What happened

04:51 ? Yes. Um, that has endpoints, that old one and no

04:58 . So there's there's two entries and it can follow them for the other

05:05 very simple and similar, and one them is that has a negative value

05:10 the other one, and it doesn't too much which one you would makes

05:15 negative value and the other one that value the main point. And it

05:22 necessarily need to be one. They be waited as well. But this

05:26 a very simple examples to show the on the matrix that's a gap.

05:32 in the terms of the mash craft the lower left hand corner and the

05:37 incidents matrix, you can kind of the happen as well. First day

05:44 one and two. They have the structure as for the line graph,

05:52 then when it comes to three, the connection with one and four So

05:57 to non zero interest are a bit apart and correspondent to again.

06:03 their role for another one in four . Now on the right hand

06:09 this Laplace and matrix that corresponds Thio . Same number off rows and columns

06:15 the number of notes and the And then you get entries, according

06:21 . The connectivity. So what you you have the center of the

06:28 it's it's themselves like which are then the dia Galo in this fashion

06:36 And then you have the off diagonal that tells toe which no,

06:41 much. No, There is an so often take note to. Then

06:46 connects both the note run and Three, so they have interest and

06:51 to not wanna know free. And this case, the elements of the

06:58 equals the number off. Incident edges that note. No one can do

07:04 same exercise for the Are there for nice craft in the bottom row on

07:15 , and it's simple, so it a lactation matrix. If one has

07:20 incidents matrix, one can get to passion by basically taking the Indians matrix

07:28 its own transports. So in this , take top one which is a

07:34 by forward, and I should take transpose over to get the four by

07:38 . And then the product is basically fight by Fine. That is several

07:42 M matrix. So it turns out this fellow Fiedler I was a check

07:51 that realize that this, uh, actually a pretty good idea for you

07:58 graph partitioning. And because of the way that the passion matrix has

08:08 , it's symmetric and Israel. Um mean that the edge values aerial values

08:16 the maybe weights and so on. it's a really honest symmetric matrix that

08:20 again values both Rio and, um or zero, but not negative,

08:26 should say. And it also has Eigen vectors correspondent to the again values

08:35 positive. I guess I should Is there anyone that Yes, For

08:42 knowledge, I'm not sure what I do about it. But anyone that

08:45 not know about idea values and Eigen . Hopefully you have heard the concept

08:59 for me, my intuitive sense is the string illustration in the couple of

09:08 backs that the I again mode for one of these. Yeah,

09:18 molds Here on the string there is Eigen value, and the deflection of

09:22 string corresponds to the idea of So those are the intuitive just notional

09:33 again, value on again vector competitions you so the point that it says

09:39 this slide. So if you have graph that turns out that the smallest

09:46 value is in fact zero because it's is disconnected cept or it doesn't has

09:56 anchors, is filthy then, lunch and sort together again values in

10:04 over their magnitude because again, they're simple to order on. Then the

10:11 non zero Eigen value corresponds to Yeah, made their crash in this

10:17 , the one that's about looks like like a sine wave on this particular

10:23 . And, um, the in case, we can use this deflections

10:32 the strain from the Bible state. the ones that have a positive and

10:38 negative deflection thio, uh, make sort of repetition ings or if you

10:44 string model in terms small elements and dynamic equation for the strength than,

10:53 , you get deflection. Correspondent of small partitions of the string is used

10:58 modeling, differential equations or other stream . So that's his son used for

11:06 petitioning. So that's what I said this slide, I think Andi So

11:14 is a little bit of analysis was by Jim Jamela at Is He Berkeley

11:22 were well known in America analyst, we don't need to get too much

11:27 again value problems yes, to, a little bit hard works. And

11:32 there are software packages out there that these things. You don't have to

11:37 to write code for it if you to use it at some point.

11:45 , so. But basically, says compute, they don't have to compute

11:50 the arguing values you basically are interested for doing a by section. The

11:55 non zero I can imagine when they in order magnitude on you. Look

12:01 the corresponding I am back to, it takes the I get back to

12:06 . Positive values to form on petition Yangon, values were non positive values

12:12 the other petition and you're in The next thing I have a little

12:20 more slides. For those of you may be interested in this in particular

12:24 the next several slides I have is show you a little bit what the

12:28 of using this bathroom might be the . In fact, kind of old

12:33 from worked at the long time ago number of collaborators. But the point

12:40 taking this showing this picture is a of things. One is first to

12:49 that the cuts or the partitions that being made intuitively makes sense,

12:55 So the algorithm didn't try to slice along the shaft of this particular

13:02 Try to minimize the cut area, ? So it, oh, make

13:06 cuts off the structure. Our particular kind of. When we talked about

13:13 inertia, a coordinate based method it's clear the moment of the

13:19 If you look at the shaft, along the shaft, and the cut

13:21 per particular to the chef. This was also done, so each

13:27 another petition should have roughly same number finite elements in this case, so

13:33 can see that color coded part if is very small elements of high density

13:41 elements, sadness in terms of the space petitions are small, but it's

13:48 the same number of elements in all petitions. Another part in using here

13:55 to show, uh, how things off scale of computational effort. Respect

14:04 the number of petitions that one would on the horizontal axis, in this

14:10 is a logarithmic access. Where's the access is just a normal in your

14:17 access. So what, You can this particular method in terms of generating

14:24 petition ISS that it, um it even grow as quickly as the log

14:33 the number of petitions. So it's computational efficient in that sense to generate

14:40 large number of partitions. And then some statistics about how many edges so

14:47 of edges have a cut to the number of edges and not show a

14:52 more examples. All this method for other structures it this one is a

14:59 simple one, um, that the partition in this way, and for

15:05 , it's kind of a curiosity because particular mhm part are petitioning was used

15:17 , you know, Foster on the of his book for Designing Colonel

15:21 So that was the work within and , I guess, like this particular

15:29 and ask difficult uses for his book . Anyway, this is another

15:33 In this case, it's kind of to linear in terms of the

15:39 in terms, off the logarithmic or of the number of petitions. Of

15:46 , the number of edges cut this . The percentage share this clearly it's

15:52 the very high percentage, but it's on the size of the partitions on

15:57 size of the graphs. It's not that particular magic number. And the

16:03 went, Yeah, again, the in the number of petitions and our

16:08 your petition there is a little bit off, perhaps more serious, large

16:15 , larger scale examples. So this for fluid dynamics, competitions that put

16:21 at the time, and in that's a curiosity. So this when

16:25 was the work was done, there the first time, um, and

16:30 aircraft was model and the airflow around aircraft waas simulated. So that was

16:38 Falcon jet that is I am regional talk chat, and you can also

16:47 some other the airliner that doesn't What? This study. But it's

16:51 a half of the triple Seven structure the affecting you may be familiar

16:56 but the point of this thing. you can see in this case,

17:00 number of notes in this final element . They went up to about 350,000

17:08 the number of elements for about two and the number of edges for about

17:12 million. And the largest graphs, , but this basically shows them a

17:18 bit off. The scalability that I to again show for these more,

17:24 somewhat ambitious or complex structures is that did this for up to about 2000

17:30 2000 petitions. And again, if look at the petition in time

17:37 uh, so it's a factor off , I guess, from the foul

17:45 Thio. The M six wing there a two million, but if you

17:52 at the time, it grows by less than a factor of five.

17:58 again, this case very well. a question. Out of curiosity,

18:03 were the processors at the time with use. So this waas um

18:14 there was on the first The computer a connection machine. I have to

18:18 out what year it was in this of the connection machine we may have

18:23 , um, but the Yes. it says number of processes 2000.

18:34 this was against the connection in model on the processors that we used on

18:48 one. I need to go That's a good question. I should

18:59 this by art, but it may been a MIPS processor. I believe

19:05 waas for this generation of the Um, Thio, whatever is your

19:18 your cell phone is more powerful than processor waas. Yeah, Um so

19:29 a good question. I will try go find out what we actually

19:36 So I just the security Oscar um, for seven years. So

19:43 early machine of this earlier version of machine 64,000 processors, which is still

19:53 fairly large number even compared to today's machines and this was done have

20:00 embarrassed to say about almost 30 years . So all right, so and

20:09 to show that was the floors. then I have your structure mechanics,

20:14 , fracture mechanics problem. And then is a case. What? I

20:20 to show for it. Weak scaling this particular case, using is,

20:27 , but for petitioning algorithm. So weak scaling in the sense that you

20:32 at the table below and tells you about the size, the number of

20:37 of the problem size on. You see if we scale the problem with

20:41 number of processors in this case. huh. You get the same efficiency

20:48 the time is proportional. The time processor is pretty much constant. So

20:56 kind of fairly ideal scaling one could at the time, people think thought

21:03 I wasn't possible for unstructured much and that was one of the things

21:08 cool people wrong with. Alright, other questions on that before ever?

21:14 are so all right. So now about multi level graft petitioning. And

21:30 a zai mentioned last time I didn't on it. This time I've been

21:34 this spectrum by section I was was today to that they only need take

21:41 about one Eigen value eso there argon methods after asked you to focus and

21:49 get one and not all of Normally, Eigen value competition is very

21:55 . But if you restrict your computations just finding one, it's not all

22:04 that it was acceptable effectively to do . And the quality of petitions were

22:10 good. And one can actually prove they are actually hasn't optically on the

22:19 sponsor you could get. If you're it up the other part, what

22:25 it feasible? One doesn't need to the Eigen values very accurately. So

22:30 our case, it was shown for the examples they tried, uh,

22:38 what customers did. So I had kinds of stuff that traded. This

22:44 was gave sort of good enough There was no really point to the

22:50 at the higher position. And that if you use destructive algorithms, then

22:55 can take advantage off. Yeah, level of precision you need to reduce

23:02 computational effort. Nevertheless, if you a very large graph, um,

23:09 computation, time may still be And in the slides that did not

23:15 about what they are in the Sign of uninterested can look at those

23:19 . Um, then you will see what fraction of the computer time is

23:24 petitioning time. But in our I think there was, yeah,

23:31 to 20% off the solution time. right, so there, there,

23:39 terms of of the level petitioning, that when the try to find a

23:46 graph that is small compared to unity small enough compared to the crafts,

23:55 want the partition. So it proceeds trying to do what's known as a

24:02 off the graph to reduce it to smaller. And then when petitions,

24:08 smaller graph by any one of the I've talked about so far and then

24:14 try to project back onto the original the partitioning that made on this course

24:24 . So here's kind of a cartoon off this multi level and thing

24:29 One goes through a number of coursing until the finance, something that I

24:35 competition and feasible to petition. And van can run the course in the

24:42 reverse to make it finally, anti anyone get back to the original

24:48 and then each step off sort of the coarsening. What makes this production

24:55 potentially refinement step four steps for each stop. So this is the basic

25:09 in this multi level partitioning, escaped preneurs kind of some cartoon examples

25:17 and then some actual results off doing . So there was one package out

25:28 . They're known as neatest. There created by faculty and students at the

25:35 of Minnesota that has become quite popular use for graph partitioning. And on

25:47 side there is pointers to Harmon's and and with this particular petition and always

25:54 the way they dio coarsening of the is to use what's known as maximum

26:01 ing's. And when maximum matching XYZ , try Thio. Find the collection

26:11 is, or the largest collection of can get, um, without anyone

26:18 the registry collect or grabbed from the sharing an end point. So if

26:25 looks like this simple graph here, the red, red and just do

26:33 share any and points and except for of the notes, you can tell

26:43 all the notes are covered by these red edges and there is one point

26:50 if you were to try to use or its incident edges has kind of

26:55 the left graph here towards the on the right side, in the

27:02 one of the three on the right of the graph, any one of

27:07 just would then connect to another note is already covered by already. So

27:13 just becomes by itself. Then, this course, a graph on basically

27:19 . They know that the ends of edge into the course note. So

27:25 this case, that becomes five course , one being the famous originally

27:32 not we did not pick an incident for in terms of the matching

27:39 but then we have a year and then one can continue to do

27:43 personally, but they didn't do it this case. So in that case

27:49 competition, that graph, uh, stops there, and then one has

27:56 do the back production onto the original . But I hope I have on

28:03 slide, yes. So launch and protected them onto the original breath

28:11 So this is kind of the idea the Minnesota folks used for the media's

28:19 , and then they used. If , as it says on the

28:25 this slide and one can use them Colonel Conlin method not to talk about

28:34 when you have in this case, by partition and then you can try

28:37 figure out is by swapping notes between . If it actually reduces the total

28:45 off edges or the weight off the is being cut. So So they

28:54 and I have some results. Something , some slides coming up,

29:01 Yes. So this as a bunch standard, um perhaps that are available

29:12 at various websites. I don't think orders, if your words websites that

29:21 interesting graphs for the community to use test algorithms. And so this comes

29:29 one or several of them. I remember. But in the paper that

29:35 called it at the bottom off the , anyone interested can find more

29:40 But what they did for this, , have somewhere they got them

29:46 As you can see in this many of them come from solid or

29:51 mechanics type applications. They're not particularly . They're not trivial but I was

30:01 they did this work. Reasonable test cases, Yeah. So a

30:13 bit more details about what's happening in maximum matching. Okay, so it

30:21 totally fully define how you take the not to include in this or construct

30:30 maximal matching set. So what? group they tried a few different schemes

30:39 on this slide. So the first is simply just randomly pick edges as

30:44 as they don't have any common, , and points or notes, and

30:52 can't doing that. So of course statistical, and you run it.

30:57 huh. A bunch of times and some statistics on die. Don't know

31:04 how many times they did it, it probably took the best all a

31:09 of runs to get the course craft doing this random matching. And I

31:23 to find it in their papers, before the lecture, and it wasn't

31:29 to me to what? How far went in terms off. They're questioning

31:36 they went all the way to 32 or the stopped at some part and

31:42 , um, partitioning before they went 32 petitions. Um now the have

31:54 successor. Heavy edge matching is the that makes logical sense to me because

32:02 they do in that case is basically visit snows in the graph. And

32:07 you want to know, you pick edge. For that note, that

32:13 not conflict in with what have you already picked that has the maximum

32:21 So that means if you pick that that becomes now a kind of course

32:27 that she, like s O, gets not the left to partitioning.

32:36 you kind of make heavy edges internal course knows. And then they did

32:43 the opposite. That was like edge . And then they also just a

32:49 bit more sophisticated than trying to find the clicks sort of heavily connected note

32:57 in the HCM notion. So what out for this particular example? Shown

33:04 this graph for 32 we're petitioning this age edge matching worked well and most

33:14 the graphs. So on the next and some of the data you picked

33:18 that, um, they used this h matching because it performed so well

33:25 the bulk of the examples they So the other part Waas then strategies

33:35 refinement when you do the UN So they tried to do the back

33:40 to the original graph on the cut that you're made on this course

33:49 And there's a whole developing number and alternatives that tried on the slide.

33:55 first is just greedy that I said one interational, this corner gone Lynn

34:05 , refining scheme that I talked about time on the KR is best that

34:11 weren't until they kind of converges. that's more than one interaction. Whatever

34:18 of integrations it took for some then the D. J R

34:25 Instead of looking at all the notes the petitions and during refinement, just

34:31 the notes that are on the edge boundary off a petition that then has

34:37 to another petition. And so that be in front of G, R

34:45 K and R are then the two one Gear and kale are but then

34:50 , considering under a notes and, , all the notes and the

34:57 And finally, the combination off using Lynn refinement for course craft. But

35:07 one gets closer to the original graph have more notes than restricted to the

35:13 and just do one interational day. on, man. So in this

35:20 , in terms off the party, find what time I can see that

35:30 winners this video are album. And one looks at the, uh,

35:39 of edges cut on, then the sonogram in refinement that uses more iterations

35:49 yeah, fine and but still stays the boundary turned out to, in

35:56 cases generate the best cut. So kind of an interesting exercises on,

36:04 there's more of it in this particular . But I just wanted to illustrate

36:11 this, um, idea works on notion of doing course inning petitioning and

36:17 course inning. And then, um has been a very successful and white

36:25 SNC. Their software has been widely for graph partitioning uh, in their

36:32 . They have done there at this . That is quite a few years

36:35 . Now the uh huh, they that it didn't matter too much for

36:43 . Petitioning method is used for the craft because it was small enough that

36:49 didn't matter whether they used something kind Lin refinement on the course craft and

36:56 around with moving stuff around or the spectral a section or other methods.

37:04 it's fairly arbitrary what you did. that case, it was small enough

37:07 it didn't matter. So neither in of running time nor in terms of

37:13 quality of the final petitioning. so any questions on this scheme on

37:30 graph partitioning? And I said, graph partitioning and well, I wouldn't

37:44 it was considered the solve problems for more than a decade. Uh,

37:54 it's kind of good enough that there that much research activity in terms of

38:01 petitioning. But in the last 5 10 years, that has picked up

38:08 . Mm, when both problems, machines has gotten significantly larger, and

38:14 come to that in a bit Now, in the early days,

38:25 , it's like a Mideast package and much all the other packages up There

38:33 are. This was sequential, so basically ran on a single note.

38:38 that means you couldn't really worked with that didn't fit on a single

38:45 So, um however, the spectral section that by myself, false

38:52 And we always did it. That's parallel partitioning methods. So we run

38:58 the few that actually did Carol So. But the media's people also

39:02 that so they now have. Oh known as far meetings for Carol Meters

39:11 again is available for download on the . But this line just that,

39:17 to point out there's a little bit to do things in parallel because if

39:23 look at the two, the top of the slide can see if each

39:29 looks at the degree off. No on the boundary here on they have

39:34 vino that has in three, the way three on the internal edge

39:40 to W. And then I was weight of five. So taking be

39:45 and putting it in partition J instead I well, then make this edge

39:51 weight five. The internal to the position, and that would reduce a

39:58 23 from the connection from I to . Now, the Day guy position

40:04 make the same conclusion that if I you are to the other I petition

40:11 the weight would only be too. this, wait five will be on

40:15 other side. Now if they don't , then they both swapped. And

40:21 that case, in fact, things worse. You can say there's another

40:27 on the bottom. It's a similar . So it's you're trying to paralyze

40:32 . Then things becomes difficult, and may not work the way you hope

40:39 . And so one thing is to what it says on this slide

40:43 Thio, bit of serialization Z or just look at petitions, pair wise

40:51 based on that. But, it's no guarantee again that it actually

40:56 better. And how many integrations you need to do in this kind of

41:02 speak. Uh, on now here's just for reference to things and

41:11 So what, in the end, the outcome in terms of the skin

41:15 was used to be the parallel And they're not entirely small, but

41:23 giant either. Still, mostly engineering problems. So here is kind of

41:30 the graph along, referring to in off how things worked out. So

41:37 is basically that the parallel is equal doesn't equally good job in terms off

41:45 number of ages cut as serial And if things are above one,

41:50 means that the parallel petition did not the as well. But in this

42:00 , when I didn't think quite as , it didn't really change things all

42:06 much You can also notice, in case, actually, the parallel petition

42:12 somewhat better. And remember that this kind of has statistical effect. So

42:19 they ran it long enough for a in there, all the cases is

42:25 what it is, Uh, not iconic and what you do but in

42:31 , And when they did this serialization , uh, the privacy got the

42:39 petitioner to generally call it cuts qualities to their serial Albelda. And as

42:53 can guess, on the horizontal axis just a bunch of different press.

43:05 , any questions on that? We're about this now. So So this

43:20 the Minnesota teams on the things that they're especially worked with the current and

43:26 algorithm and then they this multilevel abortion the corner gone Lynn and they got

43:34 . The carnival in principle is very , as I mentioned when I talked

43:40 it. In practice, it seems converge in various reiterations, and that's

43:45 kingdoms. Varying computational lee competitive Petitioner Thio the spectral reception, but in

44:00 of the quality of the cuts and two of them are comparable. And

44:05 the spectral perception gave a better cut the multi level that generally took more

44:13 . So the spectral protection folks that a group at Sandia National Labs in

44:21 that question the spectral by section They also then did the multilevel version

44:28 their code and then, But they choose to do different courses and strategy

44:35 used, once known as maximum That's when they start trying to pick

44:43 instead. And that is find it maximum set of notes that are not

44:55 , connected to each other. Exactly how the outcome differs, respected

45:02 maximum edge matching. It's not intuitively to me, but they tried

45:09 and then they tried some refinement strategies well on in terms of the what's

45:17 a surveyor, coach and reiteration to petitions anyway, Here is kind of

45:22 administration of this maximum independent old sets that illustrated just for line graph.

45:35 it shows here. The DS then A knows that cats and them together

45:40 of course, the edition in But the picking is not based on

45:45 , but based on notes. So then, of course, is in

45:50 case. That's the way that a in der Graaff on that's a little

46:00 off. What I want to come . This in terms off this basic

46:08 petitioning where the graphics models by nodes edges and you try to minimize the

46:16 of edges cut. And, as mentioned the because of the computational

46:29 is trying to come in and pretended be the most widely used because it

46:37 fairly or relatively inexpensive compared to other and generic. Quite good. Uh

46:45 . Partitions, but sometimes got in spectrum. Method did a better

46:51 but it was stolen. Um, I will talk a lot about typographic

47:04 any questions. All right, so hyper graph is trying to capture other

47:19 that the simple graph model does And here's a few examples where the

47:27 graft model turns out to work somewhat in our area. The cartoonish slides

47:38 up to illustrate how this help a models, in fact, helps sometimes

47:45 better partitions. So I know some you interested in or know about general

47:59 . Sometimes when I triangular systems and not, um so that case,

48:04 turns out that the high program model a better job in terms so conditioning

48:10 English systems are still very highly So it was just a quick slide

48:18 reference more in terms of where am can find? Well, did what

48:24 does what in terms about software and on papers? I guess.

48:37 Here's what I'm a cartoonist. Examples to illustrate the point, so I

48:45 think it's necessarily the best. But simple. So and the we have

48:53 petitioning, modeling things. That's what and edges that is on the left

48:59 side. And then there is the of the hyper graph on the right

49:02 side. So it's fun things about . And there, one way to

49:14 off tried to point out the benefits terms of clusters. So if from

49:21 to do, the petitioning and just on figuring out edge cuts and assume

49:28 for each edge being cut in this , no one needs to be a

49:33 being sent between the partitions. um, in this case, on

49:40 left hand side, if you look this great point A than it does

49:49 left neighbor in the other petitions in lower right and left hand corners.

49:55 right, on then itself in the . Upper right hand corner petition.

50:01 also has a vertical edge going So in that case, you live

50:07 well, basically, send two messages what needs to be communicated between a

50:15 it's left neighbor. And yeah, and it's south neighbor. So the

50:25 and that in fact, the hyper than more than fact lump these two

50:38 together. Someone would not need todo messages now. So that's so in

50:50 , in this case, the hyper mall. Actually, I would say

50:55 appropriately what it does. It identifies number off grid notes that has a

51:05 in the other partition. Now, of the software packages, like in

51:13 , for instance, they are clever . Thio Figure out what messages goes

51:22 the the pair off, uh, and the clusters so it would not

51:31 the naive thing or sending things for one on the knows in the in

51:37 petition. But I would look at the entire set the messages that needs

51:43 go between to compute notes and package up as one message. And then

51:52 has to be unpackaged on the other to figure out what part of the

51:59 relevant for each one off the notes the measure. That is the logical

52:04 . That is the data. So even if it's not done in the

52:11 , some are software implementations. the emerging off messages sent to you

52:19 single message. And here's some other to kind of is quickly illustrates the

52:27 how things may then be carved up terms of very regular petition. Great

52:33 . Ondas did the normal graph modeling and edges. It will get something

52:39 on the left hand side if you a hyper graph type model. If

52:44 go get something that ends up looking like on the right hand side,

52:49 in this case you do see some , and there is kind of more

52:58 of huh? The outcome A different , uh, five point sensor we

53:03 used for the a Kobe conch reiteration . If one does typical grab petition

53:12 society to grab petition on. These small examples of the difference. Is

53:18 a giant or lost in problems that be significantly better. Oh,

53:28 I have this triangular example and officials in this case, um, this

53:35 of petitioning where in the hyper graph generates the better outcome than the latest

53:44 on the part tooth is another package bond that one can download. Don't

53:50 what the afternoon comes from. The was done by. And the fact

53:56 the member that now is Georgia I think. Okay, so this

54:05 more or less same. Some more on this particular graph. Um,

54:15 any questions on this is kind of couple of lord topics and tried to

54:22 before the end of the class. go fairly quickly, given that is

54:30 your orientation and either use for And then we can talk about it

54:35 in detail or enough for an Okay, so the next thing I

54:48 get three more sub topics. Large graphs s I mentioned. So both

54:56 science and engineering problems has gotten larger machines has cut the more powerful and

55:02 so today finite element problems may have of millions of elements and and more

55:14 think about the Internet and Web And they actually have for that

55:27 uh, hyperlinks for linking between what do in behaviors as well. A

55:34 networks than man gets into crafts that have billions of notes and edges.

55:45 in that case, things aren't seriously scale. So that's as I mentioned

55:51 while ago today that that got this problem to get renewed interest in terms

55:59 methods to do, petitioning for real skin grafts on for large scale

56:06 And again, these problems are then large enough that they can't fit in

56:10 notes of money. Also parallel approaches the partitioning, for sure, and

56:19 don't have many slides on it. have a collection references for on the

56:27 that is interested in diving into it a couple of examples chilling on.

56:34 things can be done and what can achieved so well, talk about this

56:46 for showing I've coming. That seems be one of the better ones that

56:51 the right emerged in the last few that is based on so called neighborhood

57:05 . So there are offices and age . So it was. The idea

57:14 to have petitions that has then roughly number off edges on. Then

57:27 um, the edges, incidentally note up in different petitions. That means

57:38 those both those petitions need, the noble values for whatever you're going

57:45 compute. So in this particular then they replicate representation off the nodes

57:54 falls on the boundary off partitions and through the replication, that it's important

58:04 for corrections on the algorithm that the off the note in the different

58:11 our instinct. So that means when kind of replicate notes that will eventually

58:18 communication between partitions. So I'm not to talk about the different methods

58:27 But there is one example from the that is referenced in the answer.

58:36 were at the right hand corner of light and, um, the top

58:45 table gives the name of particular graphs the aliens. That's name. They

58:53 them for this table below it. , it shows the size of the

58:59 . So in this case, the craft was about 130 million notes.

59:06 about what? Fire billions edges now the middle table or the table

59:21 That shows a column for each one the graph, and then you throw

59:27 that table is different petitioning methods and one so that this was still done

59:34 . And, uh, there's an algorithm. Okay, It also shows

59:42 metes, uh, have problems for of the graphs that are the larger

59:49 on the fun looks only in recent Air that again has tried to use

59:56 or car Nitties for petitioning and seems to have scaled very well. So

60:04 people get in trouble. So the methods here, um thio scale

60:13 Okay. Andi don't crash from the grass on that bottom role on the

60:20 . It's this, uh, neighborhood that is petitioned. The edge sets

60:25 , replicate notes on boundaries on and then see there how the running

60:31 scales and the waves quite well in off the running time. So the

60:45 exchanges a little bit more expensive, I can't remember then this then city

60:50 stations. But so the running. I guess they it's not compatible or

61:03 . And I look at it. didn't spot it before. Um,

61:09 the D B H seems to be running time. On the other

61:11 on the table it mm seems to a different story home. Sorry for

61:21 . Not paying attention over there when make this slide. And then I

61:27 , um, one more method that not done again by the Sandia,

61:34 30 people that until the spectrum by method and software for it. And

61:43 not going to go into detail on one liner, but, um,

61:49 to illustrate a little bit of scalability they did get, um, so

61:56 the left graph on this slide, kind of purplish bluish nine on

62:04 That is for a Web data data set with a little bit of

62:12 three billion notes and about 128 billion . So it's a quite large graph

62:22 then the three curves below it on left hand side is for artificial.

62:30 graphs have to generate producing a few schemes from how to generate the

62:35 Um, graphs. Okay, but . Two point on the left hand

62:42 is to show a couple of things they got pretty good scalability and going

62:48 256 to 2000 notes Not perfect fill your data or proportion. It's paid

62:56 , but not all that bad. sort of one other thing. And

63:03 other thing I want to emphasize here these three artificial graphs that they use

63:10 arm at around er and Rand HD , are, you know, running

63:21 on the comment I wanted to make some things, uh, it may

63:33 easy or it's a falsehood to believe random graphs are a bad. I

63:39 it's my point. Random graphs generally not represent anything worse case, and

63:48 may not even be particularly good approximation anything really like in this case,

63:54 went their comments. They that set Israel what data set. So when

64:02 news for simplicity many times when a time finding data use something created by

64:12 ization. Somehow that is closer. our claim some average case than any

64:24 of worst case so and interpreting results using randomized data. One should be

64:34 fast, and here's a little bit a slide comparing they're extra code two

64:45 meters. In that case, they . It used, um to look

64:50 the two columns, more or less the center. On the table,

64:56 can see that they're extra pulp approach considerably faster, and also it also

65:04 to do considerably larger graphs. Then I need this environment has managed to

65:16 and two more topics, I any questions or not. So I

65:23 , I'm There's a lots of new that I have not familiarize myself all

65:30 much. So it's mostly pointers. anyone interested in large scale graft related

65:40 problems that can remodel by graphs? can say many of these methods were

65:50 methods. Uh, we're actually invented created by people are interested in large

66:01 data basis and alright dynamics craft not quickly. Here is just a example

66:12 and in terms of what may happen and engineering problems of a finite

66:18 passions and something crashing. So in case, clearly when these to

66:24 um, both potentially regulated in terms the on a tenement mesh, but

66:31 also the distribution of measure all notes need to be changed depending upon what's

66:37 . So in that case, some of dynamic Russians necessary. So one

66:45 is again like here that I may to redistribute your mission. Um,

66:51 it also happens since she did molecular or astrophysical simulations that, ah,

67:03 planets particles molecules are distributed in space over time to carry and maintain low

67:10 . More, many to most suffer between computer built on the the petition

67:19 sort of the brute force where you just to start all over every

67:22 But that's kind of expensive, so are then different forms of incremental methods

67:28 try to take advantage of that one with the relatively good partition and kind

67:37 refine it or improve it. Respect . What's happening in your application and

67:43 are the same methods can be and there's also kind of what's sometimes

67:51 as diffusion methods that they're basically looking , um, readjusting the graph by

68:02 or diffusion equations effectively how things should . And I think that's what I

68:12 here. You have some packages that be used for dynamic partition. Uh

68:19 . I don't remember if Parma disguise made this ends all time, have

68:25 visit versions for dynamic or this Just . Yes. You can use

68:31 You have a starting point remembering course and refinement, etcetera. So we

68:39 certainly use what petition? Don't already . Thio cut out the number of

68:47 steps in the partitioning. And then more thing I wanted to bring

68:56 uh, trying to finish on time that cool iss spec Titian ing.

69:05 does things go? So everything I've so far is about partitioning the data

69:14 and trying thio related to the objective , yet no balance. And,

69:25 , minimize the amount off communication. , the communication is also depending on

69:38 the partitions ends up. Respect to communication networks that into connects the notes

69:43 the cluster. It's you are interested this, and you read papers.

69:53 of most of the papers that are in the the Star references thoroughly ignore

70:04 mapping park. Some of them actually mentioned that, yes, there is

70:09 to it than yes, partitioning So it's very important. Thio not

70:18 that when you look at potentials, know benchmarks are running tests that people

70:28 report both. What? The mapping Andi. Better yet, try to

70:36 out or interpret results in terms of ? Um, I think waas So

70:43 will just make a few comments on . So here is kind of what

70:53 called. I don't think there's unnecessarily interpretation on what they are except the

71:01 two confident natural order. So partitioning . You may end up having a

71:13 D or a number label on each . And then, of course,

71:19 compute notes also have an address or I. D. And so one

71:25 is basically to have partitions Map to compute notes, according Thio, the

71:37 identities. The other one I call because when you run your partitioning

71:47 um, implicitly in how things are done, there will be a partition

71:53 each. No, it does not mean that they're ordered, according to

72:01 I labeled natural order. But they somewhere. No one of those two

72:11 to be map ings that reduce congestion the network or Leighton. See,

72:21 not necessarily clear they can be bad they can be good. The other

72:29 that is recently well define is randomly petitions to notes, the computer

72:38 and third, more sophisticated. One to do that for a map where

72:44 or partitions two notes and people have it with some resource managers. And

72:51 don't know of any one of the managers that are typically installed. In

72:57 , that does. That you may , in terms off the M.

73:02 I and that I talked about certain on the mapping was left open for

73:10 . Implementer is to potentially find out about the interconnection networks in map processes

73:17 them. So been in Eastern. , give a couple of examples.

73:24 s Oh, this is from the people. Um, there's on the

73:29 that happens to be three on the is just too different graphs that they

73:37 . And I just want to focus the collard boxes and in this

73:44 for 128 partitions. Case for the and the arbitrary mapping is exactly what

73:56 partitioning happened to result or whatever machine was run on, and they gave

74:03 one particular time to do this problem the bottom table. The corresponding 1

74:11 petitions for the autograph has two other and the random assignment that more or

74:18 gave the same as the kind of . But then it is truly randomized

74:24 . And then there's a pre partitioning you try to be clever about how

74:29 map the partitions onto the Nolde. what the interaction between partitions are

74:38 In this case, the random and ended up being pretty similar both for

74:44 auto and for the and the world graphs. And in both cases,

74:52 somewhat network career mapping did, got a job. I have something

75:02 my own experience from way back, , in which case there was kind

75:07 small problems and use and find the stuff. And in this case,

75:12 , it turned out that the random did better than whatever the default were

75:22 . To use the same word as the previous, uh, slides generated

75:28 terms off. Well, running time the communication part together is kinda operations

75:38 . There's a gun showing what the for the same machine that random medications

75:43 this case did a much better So random allocations again and not the

75:53 case it's actually tend to be a case has appointed out from the previously

76:00 what comes out of there. Whatever out of the petition may not be

76:06 that great. I think that's there's 11 more example showing in terms off

76:15 on the random mapping. Most cases have time or in the worst,

76:20 least than so. That turned up this machine. A random ization cost

76:25 congestion in the interconnection, of Yeah, and it's just in the

76:30 . Excited that. And I think what I had for today, and

76:35 time is really up. But I'll questions. So these are packages out

76:46 that I encourage you to take a at. If you are doing in

76:52 large escape problems now or in the , at least you have a reference

76:57 starting point so and stop

-
+