On 2020/7/1 5:32, Vineeth Remanan Pillai wrote: > Sixth iteration of the Core-Scheduling feature. > > Core scheduling is a feature that allows only trusted tasks to run > concurrently on cpus sharing compute resources (eg: hyperthreads on a > core). The goal is to mitigate the core-level side-channel attacks > without requiring to disable SMT (which has a significant impact on > performance in some situations). Core scheduling (as of v6) mitigates > user-space to user-space attacks and user to kernel attack when one of > the siblings enters the kernel via interrupts. It is still possible to > have a task attack the sibling thread when it enters the kernel via > syscalls. > > By default, the feature doesn't change any of the current scheduler > behavior. The user decides which tasks can run simultaneously on the > same core (for now by having them in the same tagged cgroup). When a > tag is enabled in a cgroup and a task from that cgroup is running on a > hardware thread, the scheduler ensures that only idle or trusted tasks > run on the other sibling(s). Besides security concerns, this feature > can also be beneficial for RT and performance applications where we > want to control how tasks make use of SMT dynamically. > > This iteration is mostly a cleanup of v5 except for a major feature of > pausing sibling when a cpu enters kernel via nmi/irq/softirq. Also > introducing documentation and includes minor crash fixes. > > One major cleanup was removing the hotplug support and related code. > The hotplug related crashes were not documented and the fixes piled up > over time leading to complex code. We were not able to reproduce the > crashes in the limited testing done. But if they are reroducable, we > don't want to hide them. We should document them and design better > fixes if any. > > In terms of performance, the results in this release are similar to > v5. On a x86 system with N hardware threads: > - if only N/2 hardware threads are busy, the performance is similar > between baseline, corescheduling and nosmt > - if N hardware threads are busy with N different corescheduling > groups, the impact of corescheduling is similar to nosmt > - if N hardware threads are busy and multiple active threads share the > same corescheduling cookie, they gain a performance improvement over > nosmt. > The specific performance impact depends on the workload, but for a > really busy database 12-vcpu VM (1 coresched tag) running on a 36 > hardware threads NUMA node with 96 mostly idle neighbor VMs (each in > their own coresched tag), the performance drops by 54% with > corescheduling and drops by 90% with nosmt. >
We found uperf(in cgroup) throughput drops by ~50% with corescheduling. The problem is, uperf triggered a lot of softirq and offloaded softirq service to *ksoftirqd* thread. - default, ksoftirqd thread can run with uperf on the same core, we saw 100% CPU utilization. - coresched enabled, ksoftirqd's core cookie is different from uperf, so they can't run concurrently on the same core, we saw ~15% forced idle. I guess this kind of performance drop can be replicated by other similar (a lot of softirq activities) workloads. Currently core scheduler picks cookie-match tasks for all SMT siblings, does it make sense we add a policy to allow cookie-compatible task running together? For example, if a task is trusted(set by admin), it can work with kernel thread. The difference from corescheduling disabled is that we still have user to user isolation. Thanks, -Aubrey
