Hello!
I am trying to use GSE-TDVP (global subspace expansion TDVP ) to do a real time-evolution.
Doing a subspace expansion before every time step of 1-site TDVP until the maximum bond dimension grows to the upper bound.
My Halmitonian is just an one-dimensional Bose-Hubbard model without long-range interactions, which starting from a product state.
When the length of the chain is 24, I set the upper bound to be 600 and every time step of 1-site TDVP, it costs about 28GB of memory after the reaching of upper bound.
After sweep 1: maxlinkdim=602 maxerr=0.00E+00 current_time=0.0 - 0.05im time=24.817
24.818610 seconds (2.61 M allocations: 28.250 GiB, 1.06% gc time)
tsofar = 100.0
maxlinkdim(psi) = 602
Size of psi: 34220216 bytes
But disaster occurs when I grow the size of the chain to 36 and the upper bound of bond dimension to 800, TDVP costs about 108GB of memory which results in an out-of-memory error and jobs on cluster get killed. ( near 256 GB of RAM of my cluster)
After sweep 1: maxlinkdim=819 maxerr=0.00E+00 current_time=0.0 - 0.05im time=122.393
122.396214 seconds (4.70 M allocations: 120.135 GiB, 2.65% gc time)
Free memory before GC is 44.067 GiB
Free memory after GC is 44.088 GiB
tsofar = 5.1000000000000005
maxlinkdim(psi) = 819
Size of psi: 140848968 bytes
I have used GC.gc() to free the memory after every time step of 1-site TDVP as mentioned in Large memory issue in time evolution using ITensors.jl with apply function - ITensor Julia Questions - ITensor Discourse, but it doesn’t play much role.
My julia version is 1.8.5.
So, is there any way to handle it? It bothers me for a long time.
Any help would be grateful!
if maxlinkdim(psi) < max
phis = Vector{MPS}(undef, kdim)
for j=1:kdim
prev = j == 1 ? psi : phis[j - 1]
# phis[j] = prev-im*tstep*(apply(HAL, prev;cutoff=cutoff1,method=met))
phis[j] = apply(gates,prev;cutoff=cutoff1,method=met)
normalize!(phis[j])
end
psi=ITensorTDVP.extend(psi,phis;cutoff=cutoff2)
end
phis=0
Nsite=1
@time psi = tdvp(
HAL,
-im*tstep,
psi;
nsweeps=1,
maxdim=max,
cutoff=0,
nsite=Nsite,
outputlevel=1,
normalize=true,
# solver_backend="applyexp"
)
GC.gc(true)