RFR: 8347337: ZGC: String dedups short-lived strings
Stefan Karlsson
stefank at openjdk.org
Tue Mar 11 08:58:53 UTC 2025
On Mon, 10 Mar 2025 21:40:33 GMT, Kim Barrett <kbarrett at openjdk.org> wrote:
>> When -XX:+UseStringDeduplication is turned on, ZGC requests that every single String it encounters is deduplicated. The Strings of these requests are saved in weak handles, and then processed by a separate thread. One problematic part with this is that ZGC treats these handles as strong roots for young collections. So, even if the Strings are short-lived they will be artificially kept alive until they get promoted up to the old generation.
>>
>> This creates an extreme amount of Strings and weak handles to be processed by the old collection, which can result in long major collections and allocation stalls.
>>
>> Other GCs have a filter to check for how old the Strings are before they get deduplicated.
>>
>> The proposed fix is to move the string deduplication requests to when the Strings are promoted to the old generation.
>>
>> Testing:
>>
>> * I've tested this with a small micro that showed how ZGC got extremely long major collections with string deduplication turned on.
>>
>> * SPECjbb2015 with a JVMTI agent that induces load and adds deduplicatable strings.
>>
>> * Tier1-7
>>
>> Note: I'm currently not aware of any non-artificial workload where string deduplication is an important optimization when running with Generational ZGC. If anyone knows of a workload that greatly benefits from it *AND* uses ZGC as a low-latency collector, then that would be highly interesting to look at.
>>
>> Note 2: the branch contains two changesets. In the first changeset I added a flag to be able to test and compare the old implementation with the new implementation. For the final PR I've removed that flag and the associated code as a second changeset. If we really want we could keep that flag, but given how poorly that implementation worked for Generational ZGC, I think we should just go with this new implementation.
>
> src/hotspot/share/gc/z/zStringDedup.inline.hpp line 41:
>
>> 39: // Not a String object
>> 40: return;
>> 41: }
>
> Consider using `StringDedup::is_enabled_string(obj)`, which combines `is_enabled` and `is_instance`
> into a single test.
>From what I can see, the `is_enable_string` takes a 'Klass*' and not an `oop`. So, that would require us to fetch the Klass bits, read the global `_klass_mode`, and decode the Klass*. I think I prefer the version we currently have because it optimize for the default (and most common run mode) that users are not using string dedup with ZGC.
Also, I don't think it is that important to optimize these checks now that the code is only run for promoted objects, which should be a significant lower number of objects compared to the number of objects that are visited by the marking code path.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/23965#discussion_r1988727987
More information about the hotspot-gc-dev
mailing list