Building the Right Environment to Support AI, Machine Learning and Deep Learning
|Bruce Eckel's Thinking in Java||Contents | Prev | Next|
It’s hard to believe that Java could possibly be as fast or faster than C++ .
This assertion has yet to be proven to my satisfaction. However, I’ve begun to see that many of my doubts about speed come from early implementations that were not particularly efficient so there was no model at which to point to explain how Java could be fast.
Part of the way I’ve thought about speed has come from being cloistered with the C++ model. C++ is very focused on everything happening statically, at compile time, so that the run-time image of the program is small and fast. C++ is also based directly on the C model, primarily for backwards compatibility, but sometimes simply because it worked a particular way in C so it was the easiest approach in C++. One of the most important cases is the way memory is managed in C and C++, and this has to do with one of my more fundamental assertions about why Java must be slow: in Java, all objects must be created on the heap.
In C++, creating objects on the stack is fast because when you enter a particular scope the stack pointer is moved down once to allocate storage for all the stack-based objects created in that scope, and when you leave the scope (after all the local destructors have been called) the stack pointer is moved up once. However, creating heap objects in C++ is typically much slower because it’s based on the C concept of a heap as a big pool of memory that (and this is essential) must be recycled. When you call delete in C++ the released memory leaves a hole in the heap, so when you call new, the storage allocation mechanism must go seeking to try to fit the storage for your object into any existing holes in the heap or else you’ll rapidly run out of heap storage. Searching for available pieces of memory is the reason that allocating heap storage has such a performance impact in C++, so it’s far faster to create stack-based objects.
Again, because so much of C++ is based on doing everything at compile-time, this makes sense. But in Java there are certain places where things happen more dynamically and it changes the model. When it comes to creating objects, it turns out that the garbage collector can have a significant impact on increasing the speed of object creation. This might sound a bit odd at first – that storage release affects storage allocation – but it’s the way some JVMs work and it means that allocating storage for heap objects in Java can be nearly as fast as creating storage on the stack in C++.
You can think of the C++ heap (and a slow implementation of a Java heap) as a yard where each object stakes out its own piece of turf. This real estate can become abandoned sometime later and must be reused. In some JVMs, the Java heap is quite different; it’s more like a conveyor belt that moves forward every time you allocate a new object. This means that object storage allocation is remarkably rapid. The “heap pointer” is simply moved forward into virgin territory, so it’s effectively the same as C++’s stack allocation. (Of course, there’s a little extra overhead for bookkeeping but it’s nothing like searching for storage.)
Now you might observe that the heap isn’t in fact a conveyor belt, and if you treat it that way you’ll eventually start paging memory a lot (which is a big performance hit) and later run out. The trick is that the garbage collector steps in and while it collects the garbage it compacts all the objects in the heap so that you’ve effectively moved the “heap pointer” closer to the beginning of the conveyor belt and further away from a page fault. The garbage collector rearranges things and makes it possible for the high-speed, infinite-free-heap model to be used while allocating storage.
To understand how this works, you need to get a little better idea of the way the different garbage collector (GC) schemes work. A simple but slow GC technique is reference counting. This means that each object contains a reference counter, and every time a handle is attached to an object the reference count is increased. Every time a handle goes out of scope or is set to null, the reference count is decreased. Thus, managing reference counts is a small but constant overhead that happens throughout the lifetime of your program. The garbage collector moves through the entire list of objects and when it finds one with a reference count of zero it releases that storage. The one drawback is that if objects circularly refer to each other they can have non-zero reference counts while still being garbage. Locating such self-referential groups requires significant extra work for the garbage collector. Reference counting is commonly used to explain one kind of garbage collection but it doesn’t seem to be used in any JVM implementations.
In faster schemes, garbage collection is not based on reference counting. Instead, it is based on the idea that any non-dead object must ultimately be traceable back to a handle that lives either on the stack or in static storage. The chain might go through several layers of objects. Thus, if you start in the stack and the static storage area and walk through all the handles you’ll find all the live objects. For each handle that you find, you must trace into the object that it points to and then follow all the handles in that object, tracing into the objects they point to, etc., until you’ve moved through the entire web that originated with the handle on the stack or in static storage. Each object that you move through must still be alive. Note that there is no problem with detached self-referential groups – these are simply not found, and are therefore automatically garbage.
In the approach described here, the JVM uses an adaptive garbage-collection scheme, and what it does with the live objects that it locates depends on the variant currently being used. One of these variants is stop-and-copy. This means that, for reasons that will become apparent, the program is first stopped (this is not a background collection scheme). Then, each live object that is found is copied from one heap to another, leaving behind all the garbage. In addition, as the objects are copied into the new heap they are packed end-to-end, thus compacting the new heap (and allowing new storage to simply be reeled off the end as previously described).
Of course, when an object is moved from one place to another, all handles that point at (reference) that object must be changed. The handle that comes from tracing to the object from the heap or the static storage area can be changed right away, but there can be other handles pointing to this object that will be encountered later during the “walk.” These are fixed up as they are found (you could imagine a hash table mapping old addresses to new ones).
There are two issues that make copy collectors inefficient. The first is the idea that you have two heaps and you slosh all the memory back and forth between these two separate heaps, maintaining twice as much memory as you actually need. Some JVMs deal with this by allocating the heap in chunks as needed and simply copying from one chunk to another.
The second issue is the copying. Once your program becomes stable it might be generating little or no garbage. Despite that, a copy collector will still copy all the memory from one place to another, which is wasteful. To prevent this, some JVMs detect that no new garbage is being generated and switch to a different scheme (this is the “adaptive” part). This other scheme is called mark and sweep , and it’s what Sun’s JVM uses all the time. For general use mark and sweep is fairly slow, but when you know you’re generating little or no garbage it’s fast.
Mark and sweep follows the same logic of starting from the stack and static storage and tracing through all the handles to find live objects. However, each time it finds a live object that object is marked by setting a flag in it, but the object isn’t collected yet. Only when the marking process is finished does the sweep occur. During the sweep, the dead objects are released. However, no copying happens, so if the collector chooses to compact a fragmented heap it does so by shuffling objects around.
The “stop-and-copy” refers to the idea that this type of garbage collection is not done in the background; instead, the program is stopped while the GC occurs. In the Sun literature you’ll find many references to garbage collection as a low-priority background process, but it turns out that this was a theoretical experiment that didn’t work out. In practice, the Sun garbage collector is run when memory gets low. In addition, mark-and-sweep requires that the program be stopped.
As previously mentioned, in the JVM described here memory is allocated in big blocks. If you allocate a large object, it gets its own block. Strict stop-and-copy requires copying every live object from the source heap to a new heap before you could free the old one, which translates to lots of memory. With blocks, the GC can typically use dead blocks to copy objects to as it collects. Each block has a generation count to keep track of whether it’s alive. In the normal case, only the blocks created since the last GC are compacted; all other blocks get their generation count bumped if they have been referenced from somewhere. This handles the normal case of lots of short-lived temporary objects. Periodically, a full sweep is made – large objects are still not copied (just get their generation count bumped) and blocks containing small objects are copied and compacted. The JVM monitors the efficiency of GC and if it becomes a waste of time because all objects are long-lived then it switches to mark-and-sweep. Similarly, the JVM keeps track of how successful mark-and-sweep is, and if the heap starts to become fragmented it switches back to stop-and-copy. This is where the “adaptive” part comes in, so you end up with a mouthful: “adaptive generational stop-and-copy mark-and-sweep.”
There are a number of additional speedups possible in a JVM. An especially important one involves the operation of the loader and Just-In-Time (JIT) compiler. When a class must be loaded (typically, the first time you want to create an object of that class), the .class file is located and the byte codes for that class are brought into memory. At this point, one approach is to simply JIT all the code, but this has two drawbacks: it takes a little more time, which, compounded throughout the life of the program, can add up; and it increases the size of the executable (byte codes are significantly more compact than expanded JIT code) and this might cause paging, which definitely slows down a program. An alternative approach is lazy evaluation, which means that the code is not JIT compiled until necessary. Thus, code that never gets executed might never get JIT compiled.
Because JVMs are external to browsers, you might expect that you could benefit from the speedups of some JVMs while using any browser. Unfortunately, JVMs don’t currently interoperate with different browsers. To get the benefits of a particular JVM, you must either use the browser with that JVM built in or run standalone Java applications. E