Hi Jai,

The error

    java.lang.OutOfMemoryError: Java heap space

indicates that the VM really has run out of memory. Presumably if you increased the heap size, it would actually be able to allocate that memory. You might have to add the /othervm test directive and add JVM options to require a larger heap.

The table size must be a power of two, so the largest table size that will be allocated is 1 << 30 or 1073741824 as you noted. That will take about 8GB of heap (in the no-compressed-OOP case). That's not terribly large, but we might want to check to see if there are other tests that require that much memory.

As you also noted, WeakHashMap eagerly allocates its table whereas LinkedHashMap and HashMap do not. I think this is an acceptable behavior variation. Note that we had to avoid this case in WhiteboxResizeTest:

https://github.com/openjdk/jdk/blob/master/test/jdk/java/util/HashMap/WhiteBoxResizeTest.java#L167

We might have to make similar special cases here for WHM.

I don't think we need to document this behavior difference. More precisely: this kind of implementation variation doesn't need to be specified. In the future we might change WHM to allocate lazily.

The API should accommodate extremely large values of numMappings. Even if it's larger than 1 << 30 and the table size is allocated at 1 << 30, it's still possible to add numMappings mappings without resizing. (Memory permitting, of course.) Doing so will violate the load factor invariant, and it might result in more collisions than one would like, but it should still work.

I think we just need to decide whether we want to have a test that allocates this much memory, and if so, to apply the necessary settings to make sure the JVM has enough heap.

s'marks


On 6/6/22 12:01 AM, Jaikiran Pai wrote:
In a recent enhancement we added new APIs to construct LinkedHashMap, HashMap and WeakHashMap instances as part of https://bugs.openjdk.java.net/browse/JDK-8186958.

Since we missed adding tests for that change, I have been working on adding some basic tests for this change as part of https://bugs.openjdk.java.net/browse/JDK-8285405. The draft PR is here https://github.com/openjdk/jdk/pull/9036.

It's in draft state because it has uncovered an aspect of this change that we might have to address or document for these new APIs. Specifically, the tests I added now have a test which does the equivalent of:

// numMappings = 2147483647
var w = WeakHashMap.newWeakHashMap(Integer.MAX_VALUE);

Similar tests have been added for HashMap and LinkedHashMap too, but for the sake of this discussion, I'll focus on WeakHashMap. Running this code/test runs into:

test NewWeakHashMap.testNewWeakHashMapNonNegative(2147483647): failure
java.lang.OutOfMemoryError: Java heap space
     at java.base/java.util.WeakHashMap.newTable(WeakHashMap.java:194)
     at java.base/java.util.WeakHashMap.<init>(WeakHashMap.java:221)
     at java.base/java.util.WeakHashMap.<init>(WeakHashMap.java:238)
     at java.base/java.util.WeakHashMap.newWeakHashMap(WeakHashMap.java:1363)
     at NewWeakHashMap.testNewWeakHashMapNonNegative(NewWeakHashMap.java:69)


This exception happens with only WeakHashMap. LinkedHashMap and HashMap don't show this behaviour. It appears that WeakHashMap eagerly creates an large array (of length 1073741824 in this case) in the newTable method which gets called by its constructor.

This raises a few questions about these new APIs - these APIs take an integer and the document allows positive values. So the current Integer.MAX_VALUE in theory is a valid integer value for this API. Should these APIs document what might happen when such a large numMapping is passed to it? Should that documentation be different for different classes (as seen the HashMap and LinkedHashMap behave differently as compared to WeakHashMap)? Should this "numMappings" be considered a hard value? In other words, the current documentation of this new API states:

"Creates a new, empty WeakHashMap suitable for the expected number of mappings
....
and its initial capacity is generally large enough so that the expected number
of mappings can be added without resizing the map."

The documentation doesn't seem to guarantee that the resizing won't occur. So in cases like these where the numMappings is a very large value, should the implementation(s) have logic which doesn't trigger this OOM error?

-Jaikiran


Reply via email to