- Posts: 16
Domain decomposition of CHOL-system
- denner
-
Topic Author
- Offline
- Fresh Boarder
Less
More
8 years 4 months ago #5333
by denner
Domain decomposition of CHOL-system was created by denner
I have been working on some bilayer-systems and ran into a problem when I added cholesterol to the systems. Because of a large "Estimated maximum distance required for P-LINCS: 2.626 nm" the minimum unit cell becomes very big (>3 nm) and domain decomposition fails except for very few processors. The values for two-/multi-body interactions are below the largest cut-off. I am using "lincs-order = 4".
Basically I am unsure of whether this value for the maximum distance for P-LINCS is correct (in which case my system is just too small to properly utilize the cores) or whether something else could be wrong causing the length to go up.
Hope some of you bright people are able to help - it would be very much appreciated :-)
Basically I am unsure of whether this value for the maximum distance for P-LINCS is correct (in which case my system is just too small to properly utilize the cores) or whether something else could be wrong causing the length to go up.
Hope some of you bright people are able to help - it would be very much appreciated :-)
Please Log in or Create an account to join the conversation.
- Pim
-
- Offline
- Expert Boarder
Less
More
- Posts: 105
8 years 4 months ago #5337
by Pim
Replied by Pim on topic Domain decomposition of CHOL-system
Are you using the latest cholesterol parameters?
www.cgmartini.nl/images/parameters/ITP/m...v2.0_cholesterol.itp
The description of the itp recommends you to use a somewhat higher lincs-order.
Nonetheless, this probably won't fix your problem, it shouldn't have such long constraint groups. All I can recommend is check again if you have any long bonds in your system (a molecule incorrectly split across PBC for example). If that doesn't solve it, give us your coordinates, mdp and output and perhaps we can help.
The description of the itp recommends you to use a somewhat higher lincs-order.
Nonetheless, this probably won't fix your problem, it shouldn't have such long constraint groups. All I can recommend is check again if you have any long bonds in your system (a molecule incorrectly split across PBC for example). If that doesn't solve it, give us your coordinates, mdp and output and perhaps we can help.
Please Log in or Create an account to join the conversation.
- denner
-
Topic Author
- Offline
- Fresh Boarder
Less
More
- Posts: 16
8 years 4 months ago #5338
by denner
Replied by denner on topic Domain decomposition of CHOL-system
Thanks for your reply.
I am using that .itp and I am aware that the lincs settings I refer to here are "too gentle". I will use the "right" ones when I solve the DD problem.
I get this error from a "fresh" system with a .gro file from insane.py, where no molecules are split as well as with a system which has been through energy minimization and short equilibration runs (while having this problem).
Here, www.dropbox.com/sh/lneoczb3lqslfq8/AAAwo...Ps8akLV-aL1ExEa?dl=0 I have a test system with 2x25 CHOL. The coordinates are from insane.py (not equilibrated but still reproduces the problem) and I use the martini_v2.2.itp-file.
I am using that .itp and I am aware that the lincs settings I refer to here are "too gentle". I will use the "right" ones when I solve the DD problem.
I get this error from a "fresh" system with a .gro file from insane.py, where no molecules are split as well as with a system which has been through energy minimization and short equilibration runs (while having this problem).
Here, www.dropbox.com/sh/lneoczb3lqslfq8/AAAwo...Ps8akLV-aL1ExEa?dl=0 I have a test system with 2x25 CHOL. The coordinates are from insane.py (not equilibrated but still reproduces the problem) and I use the martini_v2.2.itp-file.
Please Log in or Create an account to join the conversation.
- Pim
-
- Offline
- Expert Boarder
Less
More
- Posts: 105
8 years 4 months ago #5340
by Pim
Replied by Pim on topic Domain decomposition of CHOL-system
Hi Denner, we had a look and a value of 2.626 is actually normal (we get exactly the same value in our simulations that have been published) and it doesn't technically limit the domain decomposition. For the (somewhat unusual) example box you uploaded you indeed run into errors using the default options to mdrun as it takes the LINCS distance as the minimum cell size, but the dd size doesn't have to be bigger than the lincs distance in principle: if you use the -rcon flag to mdrun (e.g. -rcon 1.4) it overwrites the LINCS max distance and you can use domain decomposition even for your small box. However, it is not recommended to do that and in practice we don't often work with such small boxes on multiple processors, as they finish in seconds without parallelisation and don't scale very well.
Please Log in or Create an account to join the conversation.
- denner
-
Topic Author
- Offline
- Fresh Boarder
Less
More
- Posts: 16
8 years 4 months ago #5341
by denner
Replied by denner on topic Domain decomposition of CHOL-system
That sounds good - thanks. I had also convinced myself that it was ok. (One gets the number from adding the bond lengths right? And the five constrained bond lengths in CHOL adds up to exactly 2.6255 nm.)
I know that this box/system is strange, I just used it because it reproduced the problem :)
Do I understand it correctly that using the -rcon flag will "only" have efficiency effects (smaller unit cells -> communication across more cells -> slower runs)? I see that this is a GROMACS question, but I'll try to ask it here anyways.
I know that this box/system is strange, I just used it because it reproduced the problem :)
Do I understand it correctly that using the -rcon flag will "only" have efficiency effects (smaller unit cells -> communication across more cells -> slower runs)? I see that this is a GROMACS question, but I'll try to ask it here anyways.
Please Log in or Create an account to join the conversation.
- Pim
-
- Offline
- Expert Boarder
Less
More
- Posts: 105
8 years 4 months ago #5342
by Pim
Replied by Pim on topic Domain decomposition of CHOL-system
I am not sure, there must be a reason for the default setting, but I think it should be ok if you really want to do it. Perhaps stability of the simulation also suffers.
Indeed someone with more GROMACS expertise knows more...
Indeed someone with more GROMACS expertise knows more...
Please Log in or Create an account to join the conversation.
- denner
-
Topic Author
- Offline
- Fresh Boarder
Less
More
- Posts: 16
8 years 4 months ago - 8 years 4 months ago #5343
by denner
Replied by denner on topic Domain decomposition of CHOL-system
Okay, yes, you are probably right, anyways: thanks for solving my puzzle :-)
Last edit: 8 years 4 months ago by denner.
Please Log in or Create an account to join the conversation.
- mnmelo
-
- Offline
- Admin
Less
More
- Posts: 89
8 years 4 months ago #5345
by mnmelo
Replied by mnmelo on topic Domain decomposition of CHOL-system
Sorry I'm late for this discussion.
Pim and you are correct. As I understand GROMACS takes the length of four consecutive constraints and uses it to calculate the maximum length they might span (assuming 180° angles, I think). This is obviously too much for the case of cholesterol because the angles in the cholesterol rings never become colinear (far from it).
By setting -rcon you override this estimation. It's not slower than without -rcon, because if you don't specify it it becomes the estimated 2.6. If that system would run, it'd involve more communication.
Pim and you are correct. As I understand GROMACS takes the length of four consecutive constraints and uses it to calculate the maximum length they might span (assuming 180° angles, I think). This is obviously too much for the case of cholesterol because the angles in the cholesterol rings never become colinear (far from it).
By setting -rcon you override this estimation. It's not slower than without -rcon, because if you don't specify it it becomes the estimated 2.6. If that system would run, it'd involve more communication.
Please Log in or Create an account to join the conversation.
- denner
-
Topic Author
- Offline
- Fresh Boarder
Less
More
- Posts: 16
8 years 4 months ago #5347
by denner
Replied by denner on topic Domain decomposition of CHOL-system
That makes perfect sense! Should have thought about it a bit more...
Thanks for the clarification.
Thanks for the clarification.
Please Log in or Create an account to join the conversation.
Time to create page: 0.107 seconds