JVM heap size in 64 bit environment: Compressed OOP and beyond 32GB heap size

In a 32 bit there are many constraints on memory a JVM can use. The total addressable memory is just 2^32 = 4GB and then there is memory reserved for OS. Further, out of this memory available to java process, some is needed for PermGen, threads etc so the JVM can only use a much smaller memory as heapsize. See java-process-memory-jvm-heap-size for details.

Now in a 64 bit environment we do not have such constraints, since total addressable memory is more than 18 EB (2^64 = 2^4x(1024)^6 = 16 Exbibyte ~ 18x10^18 Bytes). Most of the java based enterprise applications do not need to go beyond 32 GB (besides the hardware constraints) so there is plenty of room left. However, there are some unexpected challenges. One such is increased memory usage because of expanded size of managed pointers (ordinary object pointer) so same application can take upto 1.5 times the heap on 64 bit platform after switching from a 32 bit platform.

Now managed pointers are aligned to 8byte address boundaries so managed pointers are addresses - 0, 8, 16, 24, 32, ... (ie 000000, 001000, 010000,11000,100000,...) and so on. To "compress" these pointers JVM developers used this strategy - scale down the address by a factor of 8 (ie left shift bits by 3 places) to get compressed pointers - 0, 1, 2, 3, 4,... (000,001,010,011,100,...) and so on. Using this strategy up to a 35 bits of a 64 bit address can be "compressed" into a 32 bit word (half word on 64 bit architecture). 

So using a Compressed OOP we can address 2^35 objects = 32 GB of heap space. Java SE 6u23 and later 64 bit JVM uses compressed OOP by default. However going beyond 32 GB of heap size will cause JVM to switch to regular OOP with increased size so total number of objects that can be referenced actually goes down even though heap size is increased. How big is this performance drop? This blog here shows around 48 GB was required to accommodate same number of objects as a 32 GB heap (with CompressedOOPs). So basically most java applications will see performance drop between 32 GB and ~ 48 GB!

There are other issues in going to large heap sizes like Garbage collection but that is a discussion for some other time.

Also see:

• This article has a little experiment that shows how going beyond 32 GB heap affects number of objects we can fit in the heap - Why 35GB Heap is Less Than 32GB
• About Compressed OOP on java docs - Java HotSpot™ Virtual Machine Performance Enhancements
• Wiki article by John Rose (Oracle) - CompressedOops
•  A very good explanation of Compressed OOP - Answer on stackoverflow

Java process memory: JVM heap size constraints in 32 bit environment

In theory, on a given computer system, the addressable memory is 2^b bytes where b is bit of the processor architecture. (Although there is much more to it like PAE and stuff. )

So maximum physical memory in a 32 bit is:
2^32 = 4X(1024)X(1024)X(1024) Bytes = 4096 MB = 4 GB

Out of this 4 GB some of the memory is reserved for OS, amount varies from OS to OS. So the maximum memory any process can occupy is less than 4GB. Further, in case of java process there are other components to the memory usage - JVM heap, PermGen space and native threads.

Roughly,
JVM Heap Size + PermGen Space + (Thread stack size*number of thread) < Maximum space OS allows for processes.

So there is limit on the maximum heap size you can have practically which depends on the underlying OS.

I did a few tests to check what is the result of requesting different Max heap size values on different systems. Following table lists result of different Xmx values for each of those systems:

Xmx (CentOS 32 bit)	Xmx (Ubuntu 32 bit)	Xmx (RHEL 64 bit)	Xmx (Win7 64 bit)	Xmx (Win XP 32 bit)	Xmx (SunOS 32 bit)	Result
> 4095m	> 4095m	> 4095m	> 4095m	> 4095m	> 4095m	The specified size exceeds the maximum representable size.
4095m	4095m	4095m	4095m	4095m	4095m	Incompatible minimum and maximum heap sizes specified
4030m - 4094m	4030m - 4094m	4030m - 4094m	4000m - 4094m	4000m - 4094m	4000m - 4094m	The size of the object heap + VM data exceeds the maximum representable size
2700m - 4030m	2685m - 4030m	3730m - 4030m	1500m - 4000m	1460m - 4000m	3750m - 4000m	Could not reserve enough space for object heap
< 2700m (approx)	< 2685m (approx)	< 3730m (approx)	< 1500m (approx)	< 1460m (approx)	< 3750m	JVM initialized

Little more on the results:

1. So when you try to go above 4096M (ie 4GB) the  error is "The specified size exceeds the maximum representable size." Which is expected since 4GB is maximum addressable memory.
2. At 4095M, strangely it says "Incompatible minimum and maximum heap sizes specified", looks like value 4095 is regarded same as 0. I am not sure on this, need to do more research.
3. From roughly 4000m to 4094m it says "The size of the object heap + VM data exceeds the maximum representable size" which is again understandable since as mentioned before java process has other components too so overall java process size is going beyond 4096M (the limit on addressable memory)
4. Going lower on heap, we have "Could not reserve enough space for object heap" so now we see those OS specific limits. Each OS allows a different amount of memory for a process and subtracting the non-heap part of java process we get the limit on heap size. Even the non heap part (thread stack size and limit on threads) is OS specific.

All of these tests were performed using 32 bit jdk1.7.0_04 with default PermGen space and default thread stack size and thread limits. If you tweak these heap size can be increased a little.
So in the limited test environments (with default settings) the trend is roughly as following:

• OS being 32 bit or 64 bit has very little bearing on the results.
• Server class Unix based OS (SunOS and RHEL) > Desktop Linux (Ubuntu, CentOS) > Windows Desktop (Win7, WinXP)

Need to check this on Windows Servers too but i think results will be between Windows desktop and Linux desktop values but closer to windows desktop OS.