Avi Kivity

@asias please review

I tested a 2s96c192t machine. Results are in throws/second/thread: Fedora 40 / gcc 14: 691798 Fedora 39 / gcc 14: 685662 Fedora 37 / gcc 12: 674885 Fedora 35 /...

```c++ #include #include #include #include #include #include #include using namespace std::chrono_literals; int main(int ac, char** av) { auto nr_threads = std::max(1u, std::thread::hardware_concurrency()); std::vector threads; std::latch before(nr_threads+1); std::atomic_flag stop = false;...

I don't think it's so important. We can keep it like this for a year or so, then remove it completely.

Seen again (test_concurrent_drain): https://jenkins.scylladb.com/view/nexts/job/scylla-master/job/next/6650/testReport/junit/nodetool_additional_test/TestNodetool/Tests___Sanity_Tests_RAFT___test_concurrent_drain/

Seen again: https://jenkins.scylladb.com/view/nexts/job/scylla-master/job/next/6657/testReport/junit/nodetool_additional_test/TestNodetool/Tests___Sanity_Tests___test_concurrent_restart/ ``` AssertionError: Concurrent test failed |0 |5 |10 |20 |40 |60 |80 135| ---------------------------------------------------------------------------------------------------------------------------------------------------------------- [verify_info ] []verify_status [v]erify_netstats [ve]rify_cfhistograms []verify_cfstats [v]erify_describering [ve]rify_decribecluster [concurrent_stress ] [restart ] [verify_info...

We should be solving performance problems here. dtest is for functional testing, not performance. It cannot guarantee any level of performance and anything that relies on it will eventually fail.

So, any cassandra-stress runs should be with very limited concurrency and very generous timeouts. Real stress testing of this is done by scylla-cluster-tests.

> > So, any cassandra-stress runs should be with very limited concurrency and very generous timeouts. > > Real stress testing of this is done by scylla-cluster-tests. > > And...

Do we need a cluster feature?