Re: [RFC PATCH 00/16] Core scheduling v6(Internet mail)

From: benbjiang(蒋彪)
Date: Wed Aug 05 2020 - 02:17:04 EST


Hi,

> On Aug 5, 2020, at 11:57 AM, Li, Aubrey <aubrey.li@xxxxxxxxxxxxxxx> wrote:
>
> On 2020/8/4 0:53, Joel Fernandes wrote:
>> Hi Aubrey,
>>
>> On Mon, Aug 3, 2020 at 4:23 AM Li, Aubrey <aubrey.li@xxxxxxxxxxxxxxx> wrote:
>>>
>>> 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.
>>
>> In ChromeOS we are considering all cookie-0 tasks as trusted.
>> Basically if you don't trust a task, then that is when you assign the
>> task a tag. We do this for the sandboxed processes.
>
> I have a proposal of this, by changing cpu.tag to cpu.coresched_policy,
> something like the following:
>
> +/*
> + * Core scheduling policy:
> + * - CORE_SCHED_DISABLED: core scheduling is disabled.
> + * - CORE_COOKIE_MATCH: tasks with same cookie can run
> + * on the same core concurrently.
> + * - CORE_COOKIE_TRUST: trusted task can run with kernel
> thread on the same core concurrently.
How about other OS tasks(like systemd) except kernel thread? :)

Thx.
Regards,
Jiang
> + * - CORE_COOKIE_LONELY: tasks with cookie can run only
> + * with idle thread on the same core.
> + */
> +enum coresched_policy {
> + CORE_SCHED_DISABLED,
> + CORE_SCHED_COOKIE_MATCH,
> + CORE_SCHED_COOKIE_TRUST,
> + CORE_SCHED_COOKIE_LONELY,
> +};
>
> We can set policy to CORE_COOKIE_TRUST of uperf cgroup and fix this kind
> of performance regression. Not sure if this sounds attractive?
>
>>
>> Is the uperf throughput worse with SMT+core-scheduling versus no-SMT ?
>
> This is a good question, from the data we measured by uperf,
> SMT+core-scheduling is 28.2% worse than no-SMT, :(
>
> Thanks,
> -Aubrey
>
>>
>> thanks,
>>
>> - Joel
>> PS: I am planning to write a patch behind a CONFIG option that tags
>> all processes (default untrusted) so everything gets a cookie which
>> some folks said was how they wanted (have a whitelist instead of
>> blacklist).