1. subsequence storage
- Posted by isaac <isaaca at MINDSPRING.COM> Jul 06, 1998
- 600 views
A thread a while back brought up the problem with sequence expansion: when a subsequence is appended to beyond the space buffered for it, the entire sequence and all its branced are relocated. I ran a speed check appending first to ten sequences, then to ten subsequences and the difference was startling: what took the first set .74 secs on my computer took the second set close to ten minutes! I assume a complex sequence is stored to a single block of memory. Why is this? I don't pretend to understand the inner workings of interpreters, but I would think a sequence would be stored as a single strand, with pointers representing substrings, so that {1,2,{3,4,5},6} would be stored not as: 1,2,<begin_string>,3,4,5,<end_string>,6 but as: memblock1: 1,2,<pointer_memblock2>,6 memblock2: 3,4,5 then, if I said seq[3]=seq[3]&9 then you'd only need to copy memblock2 to memblock3: 3,4,5,9 and change the pointer in memblock1, without recopying the parent sequence or any of the other subsequences. How are sequences really stored? The freedom of the sequence structure is one of my favorite features of euphoria, and I often take advantage of it with very deep sequences, but now I'm almost afraid to use them. greater efficiency here is now on the top of my wish list. In any case, I am curious as to how sequences are handled, but I understand that Robert might not want to give away all his secrets. Isaac -- first number is time for independent sequences -- second is for subseqences -- third is ratio -- currently set at 1000 repetitions; the test mentioned -- in the message was with 10000 -- with 1000 I got .05 vs. 4.27 include machine.e sequence q,w,e,r,t,y,u,i,o,p,seq atom tick,t1,t2 constant reps=1000 tick_rate(100) q={} w={} e={} r={} t={} y={} u={} i={} o={} p={} seq=repeat({},10) tick=time() for x=1 to reps do q=q&0 w=w&0 e=e&0 r=r&0 t=t&0 y=y&0 u=u&0 i=i&0 o=o&0 p=p&0 end for t1=(time()-tick) ?t1 tick=time() for x=1 to reps do seq[1]=seq[1]&0 seq[2]=seq[2]&0 seq[3]=seq[3]&0 seq[4]=seq[4]&0 seq[5]=seq[5]&0 seq[6]=seq[6]&0 seq[7]=seq[7]&0 seq[8]=seq[8]&0 seq[9]=seq[9]&0 seq[10]=seq[10]&0 end for t2=(time()-tick) ?t2 ?t2/t1
2. Re: subsequence storage
- Posted by Robert Craig <rds at EMAIL.MSN.COM> Jul 06, 1998
- 600 views
Isaac writes: > then, if I said > seq[3]=seq[3]&9 > then you'd only need to copy memblock2 to memblock3: 3,4,5,9 > and change the pointer in memblock1, without recopying the > parent sequence or any of the other subsequences. That's essentially how it's done. But think about your solution. Your time goes up as N-squared because you are copying ever-larger sequences. Euphoria highly optimizes the most common case of appending to a *simple* variable: x = append(x, v) to avoid (most of) the N-squared behavior, but for cases involving subscripts, Euphoria uses the slower, N-squared approach as in: x[i] = append(x[i], v) You can use x[i] = repeat(0, N) to avoid the need for appending, or you can use a simple variable temporarily while you are appending. This issue has come up several times on this list. Neither you nor most of the others came across the problem independently while writing a "real" program. It has mainly been pointed out by people who "heard" of the problem and then constructed an artificial program to demonstrate it. Even David Cuny, who first reported it, saw it in a Euphoria program that was *generated* in an artificial way by his Basic to Euphoria translator. It wasn't in hand-written Euphoria code. When you compare an N-squared algorithm against an order N algorithm, you can always increase N until the difference in time becomes huge. You didn't say what your value of N was. It's on my list of things to better optimize, but at the moment I'm working on other optimizations. Regards, Rob Craig Rapid Deployment Software http://members.aol.com/FilesEu/
3. subsequence storage
- Posted by Ad Rienks <Ad_Rienks at COMPUSERVE.COM> Jul 06, 1998
- 600 views
Isaac wrote: -- first number is time for independent sequences -- second is for subsequences -- third is ratio -- currently set at 1000 repetitions; the test mentioned -- in the message was with 10000 -- with 1000 I got .05 vs. 4.27 include machine.e sequence q, w, e, r, t, y, u, i, o, p, seq atom tick, t1, t2 constant reps =3D 10000 tick_rate(100) -- your code commented out --q=3D{} --w=3D{} --e=3D{} --r=3D{} --t=3D{} --y=3D{} --u=3D{} --i=3D{} --o=3D{} --p=3D{} -- --seq =3D repeat({}, 10) -- --tick =3D time() --for x =3D 1 to reps do -- q =3D q&0 -- w =3D w&0 -- e =3D e&0 -- r =3D r&0 -- t =3D t&0 -- y =3D y&0 -- u =3D u&0 -- i =3D i&0 -- o =3D o&0 -- p =3D p&0 --end for --t1 =3D (time()-tick) --?t1 -- --tick =3D time() -- for x =3D 1 to reps do -- seq[1] =3D seq[1]&0 -- seq[2] =3D seq[2]&0 -- seq[3] =3D seq[3]&0 -- seq[4] =3D seq[4]&0 -- seq[5] =3D seq[5]&0 -- seq[6] =3D seq[6]&0 -- seq[7] =3D seq[7]&0 -- seq[8] =3D seq[8]&0 -- seq[9] =3D seq[9]&0 -- seq[10] =3D seq[10]&0 -- end for --t2 =3D (time()-tick) --?t2 --?t2/t1 The way you do it is very inefficient. You should first allocate room for= all sequences and subsequences, and then fill them up. Appending and concatenating actually has to move and reallocate space, as you rightly said. But... -- now try these tests: tick =3D time() q =3D repeat({}, reps) w =3D repeat({}, reps) e =3D repeat({}, reps) r =3D repeat({}, reps) t =3D repeat({}, reps) y =3D repeat({}, reps) u =3D repeat({}, reps) i =3D repeat({}, reps) o =3D repeat({}, reps) p =3D repeat({}, reps) = for x =3D 1 to reps do q[reps] =3D 0 w[reps] =3D 0 e[reps] =3D 0 r[reps] =3D 0 t[reps] =3D 0 y[reps] =3D 0 u[reps] =3D 0 i[reps] =3D 0 o[reps] =3D 0 p[reps] =3D 0 end for = t1 =3D time() - tick ? t1 = tick =3D time() = seq =3D repeat(repeat({}, reps), 10) -- this is done very fast, *if* you've got enough memory! = for x =3D 1 to 10 do for z =3D 1 to reps do seq[x][z] =3D 0 end for end for t2 =3D time() - tick ?t2 while get_key() =3D -1 do end while Sincerely, Ad Rienks writing at 21:42 , = on Monday 6 July 1998 Using EMail Assist for WinCIM
4. Re: subsequence storage
- Posted by Hawke <mdeland at NWINFO.NET> Jul 06, 1998
- 563 views
snipsnipsnipcutcutcutsnipsnip Ad Rienks wrote: > The way you do it is very inefficient. > You should first allocate room for > all sequences and subsequences, and > then fill them up. Appending and > concatenating actually has to move > and reallocate space, as you rightly > said. But... > -- now try these tests: snipslicecutsnipsnipcutseversnip I ran the tests Ad gave us to try and on my machine (Pentium) the subsequence code (part2) ***BEAT*** the sequence code (part1) every time by a factor of 150% (0.06-part1 vs 0.04-part2) I found this interesting :) --Hawke'
5. Re: subsequence storage
- Posted by isaac <isaaca at MINDSPRING.COM> Jul 07, 1998
- 578 views
Preallocation via repeat is a good solution, thanks. >It has mainly been pointed out by people who >"heard" of the problem and then constructed >an artificial program to demonstrate it. Actually, I have run across this problem in a "real" program, though I didn't know it at the time. The program dealt with several little creatures scurrying around in a virtual environment, 500*500*3 4 byte spaces stored in memory. Each creature's behavior was controled by a function into which I fed a sequence repesenting the 101*101*3 area around it (50 spaces in each direction). Each 4 byte space was translated into 6-8 integers which were appended onto the subseqence representing that space. I had abandoned the project because it was just too slow, but after a few changes to the code to avoid appending, it's beginning to look viable. thanks, Isaac
6. Re: subsequence storage
- Posted by Ad Rienks <Ad_Rienks at COMPUSERVE.COM> Jul 07, 1998
- 587 views
Hawke wrote: >I ran the tests Ad gave us to try and on my >machine (Pentium) the subsequence code (part2) > ***BEAT*** >the sequence code (part1) every time by a >factor of 150% (0.06-part1 vs 0.04-part2) >I found this interesting :) >--Hawke' On my machine (P75, 8Mb) the results were even more interesting: sometime= s the first gave 0.11 and the second 0.16-0.17, but another time it was jus= t the opposite (0.17 vs 0.11). Only when I set reps to 100.000, the second test hung my machine under DOS, and gave 'data too big for memory' or something like that, when run in W95, using D. Cuny's editor. Have to tes= t that again after installing 32 extra Mb's. I've got them here, actually! Ad