ADOL-C thread safety

[chrhansk]

I am working on a project which involves the evaluation of several functions and their derivatives. At each step, the derivatives of several different functions have to be evaluated (rather than several Jacobian rows or Taylor coefficients).

I use ADOL-C to compute these derivatives and am generally pleased with the results. However, I would like to compute all required derivatives in parallel. Unfortunately, it seems that ADOL-C is not by default thread-save. In my case, OpenMP-based parallel evaluation is not an option.

The thread unsafety seems (at least in part) due to the fact that the tape data structures are stored as global variables:

ADOL-C/ADOL-C/src/tape_handling.cpp

Lines 337 to 348 in e1fe476

	vector<TapeInfos *> ADOLC_TAPE_INFOS_BUFFER_DECL;
	/* stack of pointers to tape infos
	* represents the order of tape usage when doing nested taping */
	stack<TapeInfos *> ADOLC_TAPE_STACK_DECL;
	/* the main tape info buffer and its fallback */
	TapeInfos ADOLC_CURRENT_TAPE_INFOS_DECL;
	TapeInfos ADOLC_CURRENT_TAPE_INFOS_FALLBACK_DECL;
	/* global tapeing variables */
	GlobalTapeVars ADOLC_GLOBAL_TAPE_VARS_DECL;

As far as I can tell, the manual does not mention thread-safety, so this is certainly not a bug. What I would like to have is a guarantee of the form

Evaluation sequences for different tapes are save to execute in parallel

where an evaluation sequence involves a taping step using trace_on(id), ...,trace_off() followed by any gradient(id,...) or more complicated sweeping sequences.

As far as I can tell, most of the actual work is conducted on the current tape and the set of all tapes is only touched very infrequently. So I think making the current tape thread_local and synchronizing the access to the other global variables would go a long way towards thread safety without sacrificing much in terms of performance.

I would be willing to try my hand at implementing this, provided you would merge such a feature.

[awalther1]

To the best of our knowledge ADOL-C should be thread-safe, however this feature is definitely not widely tested. Therefore I would appreciate your effort if you would fix this and I will definitely merge such a contribution to the master branch. Furthermore, I will forward your message to somebody who is looking into the parallelization of ADOL-C

Best regards

Andrea Walther

[mflehmig]

Hi @chrhansk,

ADOL-C can handle OpenMP parallel primal functions as described in [1] and [2]. For that, an isolation of global and thread-local memory is enforced using OpenMP techniques.

What parallelization API will you use?

Best regards
Martin

[1] Parallel derivative computation using ADOL-C, Andreas Kowarz, Andrea Walther, 2008
[2] Section Advanced algorithmic differentiation of OpenMP parallel programs in ADOL-C Manual

[cgraeser]

@chrhansk Have a look at #62 which introduced a threadsafe traceless forward mode.