OpenEye deems only the following toolkits “thread-safe”:
- OEChem TK (including oeplatform, oesystem, oechem, and oebio)
- OEDepict TK
- OEDocking TK
- Grapheme TK
- GraphSim TK
- Lexichem TK
- Omega TK
Absolutely no guarantee is made about the thread-safety of any other library not on the preceding list.
The most common reason to multi-thread a program is for performance. However, threading has certain overhead costs associated with it. The user should profile their application to determine the best possible way to parallelize it. There is no guarantee that multi-threading a program will make it run faster, in fact, it can make it run dramatically more slowly. Some of the examples presented in this chapter will perform more slowly than their single-threaded counterparts. They were chosen for the simplicity of demonstrating certain concepts. The user should perform timing analysis on the code in this chapter on their own systems to determine whether the methods are appropriate for them.
By default, it is not safe to pass instances of the following OpenEye Toolkit objects between threads:
Doing so will likely cause a crash due to OpenEye Toolkits use of thread-local memory caches that can not be shared across threads. If a multi-threaded OpenEye Toolkit program crashes, a good first solution to try is inserting the following at the beginning of main.
However, there is a performance cost to turning off all OpenEye Toolkits memory caches using the above. It may be possible to gain that performance back by using an allocator better tuned to the specific application. For multi-threaded applications, the following allocators are known to be very performant:
Prior to the 2015.Oct release, molecule and substructure search objects could not be shared between threads. This could easily cause crashes to occur whenever using the following classes inside an OEOnce construct:
The above objects can now be safely passed between threads and used in OEOnce function local statics. As of 2015.Oct, the above objects now manage their own memory pools on a per-object basis. This can possibly lead to more memory waste, as each molecule has enough space to store 64 atoms and bonds by default, even if it is empty. This is why the OEMolBaseType_OEMiniMol molecule implementation uses the older memory pooling, as that is designed for memory efficiency.
OEChem TK provides a rather advanced mechanism for spawning a thread to handle I/O asynchronously. Parallelizing I/O operations can yield significant performance gains when dealing with bulky file formats like SDF and MOL2, or when dealing with files on network mounted directories. Naturally, performance gains may vary based upon the system configuration in question.
The input and output threads are encapsulated inside the oemolthread abstraction. oemolthreads are designed to behave as similarly as possible to oemolstreams. This should allow easy drop in replacement to existing code by simply changing the type of a variable from oemolistream to oemolithread for input, and oemolostream to oemolothread for output.
Listing 1: Using Threaded Input and Output
from openeye import oechem import sys ss = oechem.OESubSearch(sys.argv) ifs = oechem.oemolithread(sys.argv) ofs = oechem.oemolothread(sys.argv) for mol in ifs.GetOEGraphMols(): oechem.OEPrepareSearch(mol, ss) if ss.SingleMatch(mol): oechem.OEWriteMolecule(ofs, mol)
The difference in Listing 1 may be a little difficult to spot for the experienced OEChem TK programmer used to seeing oemolstream declarations, however, the following two declarations are not oemolstreams. They are oemolthreads meant to behave as closely as possible to oemolstreams.
ifs = oechem.oemolithread(sys.argv) ofs = oechem.oemolothread(sys.argv)
These do not make much sense in a multi-threaded environment where the I/O thread can be at any arbitrary location in the file. Using the methods would be prone to race conditions.
The default constructor of oemolistream will open the stream on stdin. However, read operations are handled by a separate opaque I/O thread inside the oemolithread. The I/O thread uses blocking reads that do not return unless the user closes stdin of the process. oemolithreads can still read from stdin using the same syntactic trick described in the Command Line Format Control section supplied to open or the constructor which takes a string.
As will be described later, the I/O threads inside oemolthreads do a minimal amount of work to maximize scalability in parallel applications. The work described in the Input and Output section to recover multi-conformer molecules from single-conformer file formats can be quite expensive. Note, this does not preclude oemolithreads from reading multi-conformer molecules from file formats that support them inherently, e.g., multi-conformer molecules can be read out of multi-conformer OEB files through oemolithreads.
With all the preceding considerations taken into account the real gain of oemolthreads is that OEReadMolecule and OEWriteMolecule can be safely called from multiple threads onto the same oemolthread. They are guaranteed to be free of race conditions when applied to oemolthreads. Not only that, the expensive parsing (OEReadMolecule) and serialization (OEWriteMolecule) occurs in the thread that calls the function, not in the oemolthread itself. The oemolthread takes care of I/O and compression if reading or writing to a .gz file.
Figure: oemolithread and Figure: oemolothread shows how data flows between an arbitrary number of worker threads and the I/O thread. Molecule data passes through a protected bounded buffer. It is “protected” against race conditions when multiple threads operate on the buffer. It is “bounded” in size to prevent increasing the required amount of memory by inserting many items into the buffer faster than can be consumed on the other side. The I/O threads will be put to sleep, allowing other threads to use the CPU, when there is not a sufficient amount of work for them to do. This means the number of worker threads should typically be the number of available CPUs. The I/O threads don’t require much CPU time, therefore, they don’t need their own dedicated CPU. This could change as the number of worker threads increases, or if the I/O thread is expected to handle data compression or uncompression.
Listing 2 demonstrates how to scale the molecule searching program to multiple CPUs. A worker thread is dedicated for each CPU. There are also I/O threads to handle the input and output operations. The data parsing (OEReadMolecule), substructure searching(OESubSearch.SingleMatch), and data serialization (OEWriteMolecule) is all done in parallel by the worker threads. The costly synchronization point of reading and writing data is handled by the separate I/O threads.
Listing 2: Using multiple threads to read and write to molthreads
from openeye import oechem from threading import Thread import sys ss = oechem.OESubSearch(sys.argv) ifs = oechem.oemolithread(sys.argv) ofs = oechem.oemolothread(sys.argv) class MultiCoreMolFind(Thread): def run(self): for mol in ifs.GetOEGraphMols(): oechem.OEPrepareSearch(mol, ss) if ss.SingleMatch(mol): oechem.OEWriteMolecule(ofs, mol) thrds =  for i in range(oechem.OEGetNumProcessors()): thrd = MultiCoreMolFind() thrd.start() thrds.append(thrd) for thrd in thrds: thrd.join()
The global interpreter lock can eliminate any chance of scaling across multiple processors. It depends how much of the run time is in C++ and how much is in Python.
Listing 2 will generate the same set of molecules in the output file as Listing 1. However, there is big difference between the output files, the molecule order is non-deterministic. For most users this does not matter, database files are treated as arbitrarily ordered sets of molecules. If the order of the molecule database is important do not use the program in Listing 2.
Database order is preserved by the program in Listing 1 because there is only one worker thread. This allows users with database order restrictions to still take advantage of threaded asynchronous I/O.
OEChem TK is considered to be a “thread-safe” library. It makes the same guarantees the C++ Standard Template Library (STL) makes in concern of threads:
Unless otherwise specified, standard library classes may safely be instantiated from multiple threads and standard library functions are reentrant, but non-const use of objects of standard library types is not safe if shared between threads. Use of non-constness to determine thread-safety requirements ensures consistent thread-safety specifications without having to add additional wording to each and every standard library type.
Unless otherwise stated in the API documentation, OEChem objects can be instantiated and used in multiple threads. For example, in Listing 2, special care was taken to make the molecule object local to each thread.
This was accomplished by calling the OEChem::oemolithread::GetOEGraphMols method in each thread. An OEGraphMol is constructed local to each OEChem::oemolithread::GetOEGraphMols generator invocation.
def run(self): for mol in ifs.GetOEGraphMols():
The same object can be accessed from multiple threads if all access is through a const method. A const method is any method that does not change data internal to the class. In C++, a method can be explicitly marked const, allowing the compiler to cause the code to not compile if this guarantee is not kept.
For example, the OESubSearch.SingleMatch method is a const method. Therefore, the same OESubSearch object can be used from multiple threads without the need to make local copies of the OESubSearch object in Listing 2. However, notice that the molecule argument passed to OESubSearch.SingleMatch is local to each thread. Otherwise, one thread could be reading data into the molecule object while another thread is trying to perform the substructure search, leading to a race condition and undefined behavior.
OEChem functions, unless otherwise stated in the API documentation, can be safely called from multiple threads. For example, OEGetTag needs to access a global table to make string to unsigned integer associations global to the program. OEGetTag ensures its own thread safety so that it is safe to call from multiple threads.
However, mutual exclusion can be very expensive. This is why OEChem TK (and the STL) make no effort to make concurrent object access safe. So when mutual exclusion can not be avoided, as in the case of OEGetTag, faster alternatives should be sought. For example, the return of OEGetTag can be cached to a local variable.
It is important to realize that even though the function may be thread-safe, the function may be altering the state of its arguments, which could be shared across threads. Again, the “const-ness” of the arguments must be taken into account. For example, it is not safe to call OEWriteMolecule on the same molecule from separate threads (even on different oemolostreams) because the molecule argument to OEWriteMolecule is non-const. The molecule argument is non-const for performance, the molecule may need to be changed to conform to the file format being written to. OEChem TK provides OEWriteConstMolecule that will make a local copy of the molecule before making necessary changes for the file format being requested.
Exceptions to the const argument rule must be explicitly stated in the API documentation. For example, the OEWriteMolecule function applied to the non-const argument of oemolothread is thread-safe only by virtue of design.
Memory allocation is often a bottleneck in any program. This is especially the case in multi-threaded programs because memory allocation must protect a globally shared resource, the memory space of the process. Most malloc implementations use mutual exclusion to make themselves thread-safe. However, mutual exclusion is expensive, and cuts down on the scalability of the multi-threaded program. OEChem TK attempts to alleviate the problem by implementing a memory cache for the following small objects:
The cache is totally opaque to the normal OEChem TK user. However, it can cause problems in multi-threaded situations.
The small object cache implementation is controlled through the OESetMemPoolMode function. The parameter passed to this function determines what sort of cache implementation to use, the possible options are a set of values bit-wise or’d from the OEMemPoolMode constant namespace. The possible combinations are as follows:
|Makes no attempt at thread safety. Will only cache allocations up to a certain size, then further allocations will be sent to the global operator new(size_t) function.|
|Makes no attempt at thread safety. Will cache all allocations up to any size, never returning memory to the operating system after it has been requested.|
|Protects a single global cache with a mutex. Will only cache allocations up to a certain size, then further allocations will be sent to the global operator new(size_t) function.|
|Protects a single global cache with a mutex. Will cache all allocations up to any size, never returning memory to the operating system after it has been requested.|
|Uses a separate cache for each thread. Will only cache allocations up to a certain size, then further allocations will be sent to the global operator new(size_t) function.|
|Uses a separate cache for each thread. Will cache all allocations up to any size, never returning memory to the operating system after it has been requested.|
|System:||Forwards all allocations to the global operator new(size_t) function. This allows the user to circumvent the entire caching system in preference to their own operator new(size_t) function allowing other third party memory allocators to be linked against.|
The default implementation is ThreadLocal|UnboundedCache for performance in single threaded programs and thread safety in the most common multi-threaded programs (e.g. Listing 2).
If the program is known to only have a single thread there is a potential performance benefit in using the OEMemPoolMode_SingleThreaded option. This flag tells the cache implementation it does not have to worry about being thread-safe.
Setting single threaded memory pool mode
from openeye import oechem import sys oechem.OESetMemPoolMode(oechem.OEMemPoolMode_SingleThreaded | oechem.OEMemPoolMode_UnboundedCache) ss = oechem.OESubSearch(sys.argv) ifs = oechem.oemolithread(sys.argv) ofs = oechem.oemolothread(sys.argv) for mol in ifs.GetOEGraphMols(): oechem.OEPrepareSearch(mol, ss) if ss.SingleMatch(mol): oechem.OEWriteMolecule(ofs, mol)
oemolthreads can still be used since they do not use the small object cache.
The preceding program demonstrates how to use OESetMemPoolMode to tell OEChem TK the program only contains a single thread using OEMemPoolMode_SingleThreaded. OEMemPoolMode_UnboundedCache is also used to get the maximum amount of performance out of OEChem TK. The options for the cache implementation are bitwise or’d together as shown in the following code snippet:
oechem.OESetMemPoolMode(oechem.OEMemPoolMode_SingleThreaded | oechem.OEMemPoolMode_UnboundedCache)
The default thread safety level in OEChem TK is to use OEMemPoolMode_ThreadLocal since it can offer performance nearly as good as OEMemPoolMode_SingleThreaded. It also means the vast majority of multi-threaded OEChem TK programs will be thread-safe. This is because each thread gets its own dedicated small object cache.
However, this introduces a memory leak when certain cached objects are created in one thread and then passed to another thread to be deleted.
If database order is a necessity it may be possible to gain performance from threading using “pipeline” parallelism. Listing 3 demonstrates how the molecule searching problem can be broken down into 5 stages, each running in a separate thread. In order, the 5 stages are:
Listing 3: Setting mutexed memory pool mode
from openeye import oechem from threading import Thread try: from queue import Queue except ImportError: from Queue import Queue import sys ss = oechem.OESubSearch(sys.argv) ifs = oechem.oemolithread(sys.argv) ofs = oechem.oemolothread(sys.argv) iqueue = Queue(1024) oqueue = Queue(1024) class ParseThread(Thread): def run(self): mol = oechem.OEGraphMol() while oechem.OEReadMolecule(ifs, mol): iqueue.put(mol, block=True) mol = oechem.OEGraphMol() # signal SearchThread to die iqueue.put(None, block=True) class SearchThread(Thread): def run(self): mol = iqueue.get(block=True) while mol is not None: oechem.OEPrepareSearch(mol, ss) if ss.SingleMatch(mol): oqueue.put(mol, block=True) mol = iqueue.get(block=True) # signal main thread to die oqueue.put(None, block=True) pthrd = ParseThread() pthrd.start() sthrd = SearchThread() sthrd.start() # main thread will handle output mol = oqueue.get(block=True) while mol is not None: oechem.OEWriteMolecule(ofs, mol) mol = oqueue.get(block=True)
Listing 3 shows how a molecule can be passed down a pipeline of threads.
Passing OEMemPoolMode_System to OESetMemPoolMode tells the OpenEye Toolkits to using the system defined operator new(size_t). The implementation is thread-safe on all modern systems and may even be faster than using a mutexed cache. It is up to the user to figure out which is best for their system. The following code snippet will bypass the cache entirely in favor of calling the systems memory allocation routines:
The real benefit of specifying OEMemPoolMode_System is the ability to plug in third party memory allocators. The following memory allocators are recommended for multi-threaded programs:
Two types of small object caches can be set with OESetMemPoolMode: Bounded and Unbounded. The default is an Unbounded cache for performance. The assumption is most OEChem TK programs will have a constant memory usage. Even if the program has a steadily rising memory footprint the program will usually exit after it has maxed out its memory requirements.
This may not be the case for long running programs, for example, a web service. These programs may need to create a temporarily large memory footprint and then release it back to the system for some other program to use. An Unbounded cache will hold onto all memory it sees, never returning any to the operating system.
A Bounded cache will only cache allocations up to a limit. After that limit has been reached all further allocations will be sent to the system memory allocator. This allows the deallocations later in the program to return the memory to the operating system. This allows certain objects to still use the fast cache while the other objects that are bloating the memory footprint will incur the cost of asking the operating system for the memory they require.