Hi all,
from some performance tests we learned that a bottleneck for Hibernate
using applications is often identified in the amount of memory we
allocate at runtime, even considering the so called "short lived"
objects which usually are not a threat are actually too high.
Specifically, the highest consumer is the wild high amount of
instances of _org.hibernate.engine.spi.EntityKey_.
I think we have mainly two different strategies to attack this:
1# reduce the amount of EntityKey instances being allocated
2# reduce the size of each EntityKey instance
While 1# seems a wise move, I'll leave that to the ORM experts to
consider for future as it probably requires a broader knowledge of how
all components fit together.
So I'm attacking 2#, especially as I thought I could get some high win
in short time :)
To properly estimate the runtime size of each instance, I could simply
use the various reference tables of how much each pointer and
primitive takes, but there is actually some more complexity related to
the order in which the fields will be organized, and the requirement
of object alignment.
The rules aren't too hard to figure out the cost of a small value
object, but I'm actually using a tool which nicely dumps out a
reasonable estimate: "java object layout".
So this is the output of the current EntityKey implementation:
Java HotSpot(TM) 64-Bit Server VM (build 24.45-b08, mixed mode)
Running 64-bit HotSpot VM.
Using compressed references with 3-bit shift.
Objects are 8 bytes aligned.
org.hibernate.engine.spi.EntityKey
offset size type description
0 12 (assumed to be the object
header + first field alignment)
12 4 int EntityKey.hashCode
16 1 boolean EntityKey.isBatchLoadable
17 3 (alignment/padding gap)
20 4 Serializable EntityKey.identifier
24 4 String EntityKey.entityName
28 4 String EntityKey.rootEntityName
32 4 String EntityKey.tenantId
36 4 Type EntityKey.identifierType
40 4 SessionFactoryImplementor EntityKey.factory
44 4 (loss due to the next object alignment)
48 (object boundary, size estimate)
So that's 48 bytes per instance, and matches my expectations.
In our benchmark, we created about 200GB of such instances in 20
minutes, and the impact is of approximately 10% of the overall memory
pressure of the application.
I have an open pull request [2] which improves things, changing the
code to the following layout:
org.hibernate.engine.spi.EntityKey
offset size type description
0 12 (assumed to be the object header +
first field alignment)
12 4 int EntityKey.hashCode
16 4 Serializable EntityKey.identifier
20 4 String EntityKey.tenantId
24 4 EntityEssentials EntityKey.persister
28 4 (loss due to the next object alignment)
32 (object boundary, size estimate)
So we get from 48 to 32 bytes per instance; assuming the same
allocation that means saving about 33%, so some 60GB of memory
bandwith saved.
I think we could take it as a first step in the right direction, still
we're so very close to actually improve of 50%: there are 4 bytes per
instance "wasted" just because of alignment needs, if we could remove
just one more field we would save 8bytes.
In this example I removed the _tenantId_:
org.hibernate.engine.spi.EntityKey
offset size type description
0 12 (assumed to be the object header +
first field alignment)
12 4 int EntityKey.hashCode
16 4 Serializable EntityKey.identifier
20 4 EntityEssentials EntityKey.persister
24 (object boundary, size estimate)
So I would aim at implementing this, or alternatively if we really
can't remove the tenantId we could remove the (cached) hashCode:
org.hibernate.engine.spi.EntityKey
offset size type description
0 12 (assumed to be the object header +
first field alignment)
12 4 Serializable EntityKey.identifier
16 4 String EntityKey.tenantId
20 4 EntityEssentials EntityKey.persister
24 (object boundary, size estimate)
But that hashCode calculation is really hot so I would love to rather
remove the _tenantId_.
How can we remove the tenantId ?
I was thinking to have the tenandId included in the persister, so we'd
have a new interface:
public interface TenantLocalEntityPersister extends EntityPersister {
String getTenantId();
}
Implementors would have two fields:
EntityPersister ep; //shared across all tenants
String tenantId;
Then one single instance of such a TenantLocalEntityPersister would be
shared across all EntityKey instances associated to it, but also by
many other objects which need to shrink: for example
org.hibernate.engine.spi.EntityEntry needs to access the same concept,
and is the second highest consumer of memory (so next in line for a
similar refactoring to be done).
Would that be doable to refactor EntityPersister to have this concept
of per-tenant instance?
As a nice side-effect, I suspect it would also bring the cost of
managing multi-tenancy to zero for those applications which aren't
doing any multi-tenancy, while now there seems to be a cost paid by
all users.
Sanne
[1] https://github.com/jbaruch/java-object-layout
[2] https://github.com/hibernate/hibernate-orm/pull/633
_______________________________________________
hibernate-dev mailing list
hibernate-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/hibernate-dev
