Cell splitting in neural networks extends strong scaling (Hines et al. 2008)

 Download zip file   Auto-launch 
Help downloading and running models
Accession:97917
Neuron tree topology equations can be split into two subtrees and solved on different processors with no change in accuracy, stability, or computational effort; communication costs involve only sending and receiving two double precision values by each subtree at each time step. Application of the cell splitting method to two published network models exhibits good runtime scaling on twice as many processors as could be effectively used with whole-cell balancing.
Reference:
1 . Hines ML, Eichner H, Schürmann F (2008) Neuron splitting in compute-bound parallel network simulations enables runtime scaling with twice as many processors. J Comput Neurosci 25:203-10 [PubMed]
Citations  Citation Browser
Model Information (Click on a link to find other models with that property)
Model Type: Realistic Network;
Brain Region(s)/Organism: Generic;
Cell Type(s):
Channel(s):
Gap Junctions:
Receptor(s):
Gene(s):
Transmitter(s):
Simulation Environment: NEURON;
Model Concept(s): Methods;
Implementer(s): Hines, Michael [Michael.Hines at Yale.edu];
/
splitcell
nrntraub
hoc
balcomp.hoc *
binfo.hoc *
defvar.hoc *
karkar.hoc
lbcreate.hoc *
loadbal.hoc *
mscreate.hoc
msdiv.hoc *
parlib.hoc
parlib2.hoc
traubcon.hoc *
traubcon_net.hoc *
                            
proc load_balanced_create() {local host, gid, usx, sx, si, sb, c, cc, ptree \
  localobj cell, sr
	host = load_balance_.host.x[$1]
	gid = load_balance_.gid.x[$1]
	sx = load_balance_.splitx.x[$1]
	usx = load_balance_.unsplitx.x[$1]
	si = load_balance_.spliti.x[$1]
	sb = load_balance_.splitb.x[$1]

	execute($s2)
	cell = cells.object(cells.count-1)

	if (sx < 0) { // entirely on this cpu
		pc.set_gid2node(gid, pc.id)
		gidvec.append(gid)
	}else{
		c = load_balance_.cell_complexity(cell)
		load_balance_.compute_roots()
		sr = ldbal_reconnect(si, abs(sb))
// following tests must be carried out with point process complexities in mcomplex.dat = 0
//sr.sec printf("%d a split %s %g %g    %g %g   %d %d %d %d\n", pc.id, secname(), usx, c, sx, usx - sx, host, gid, si, sb)
		ptree = 0 if (sb < 0) { ptree = 1 }
		ldbal_split(sr, host, gid, ptree, cell)
//c = load_balance_.cell_complexity(cell)
//if (host == pc.id) { cc = sx } else { cc = usx - sx }
//ptree = 0 if (!pc.gid_exists(gid)) ptree = 1
//printf("%d split %s %d %d %g %g\n", pc.id, cell, gid, ptree, c, cc)
	}
}

//reconnect so a split at the returned
// SectionRef corresponds to the complexity desired
// ptree = 1 means the parent tree will go on the host
obfunc ldbal_reconnect() {local i  localobj sp, bs, sr, sc
	ip = load_balance_.parent_vec_.x[$1]
	if (ip  == -1) {
		execerror("cannot split at original root", "")
	}
	sp = load_balance_.secref(ip)
	// the branch set (in the proper order)
	bs = load_balance_.parent_vec_.indvwhere("==", ip)
	// disconnect all the children
	// assume all children effectively connected at their 0 end to the trueparent
	// at the 1 end
	for i=0, bs.size-1 {
		sr = load_balance_.secref(bs.x[i])
		sr.sec { disconnect() }
	}
	if ($2 <= bs.size) { // is it an individual?
		// then all get connected to the trueparent and return
		// the individual
		for i=0, bs.size-1 {
			sr = load_balance_.secref(bs.x[i])
			sp.sec connect sr.sec (0), 1
		}
		sc = load_balance_.secref(bs.x[$2-1])
	}else{ // it is a sum
		sc = load_balance_.secref(bs.x[0])
		sp.sec connect sc.sec (0), 1
		n = $2 - bs.size // 1 means 0+1
		for i=1, n { // connect to the 0 child
			sr = load_balance_.secref(bs.x[i])
			sc.sec connect sr.sec (0), 0
		}
		for i=n+1, bs.size-1 { // connect to the trueparent
			sr = load_balance_.secref(bs.x[i])
			sp.sec connect sr.sec (0), 1
		}
	}
	return sc
}

proc ldbal_split() {
	if ($4 == 0) {
		$o1.sec pnm.splitcell($2, $2+1)
	}else{
		$o1.sec pnm.splitcell($2+1, $2)
	}
	if (section_exists("comp", $o5.presyn_comp ,$o5)) {
		pc.set_gid2node($3, pc.id)
		gidvec.append($3)
	}else{
		pc.set_gid2node($3 + splitbit, pc.id)
		gidvec.append($3 + splitbit)
	}
}