July 2023 Linux kernel Meetup - Call for Proposal

[Santosh Shukla]

Hi Bangalore Linux Kernel Community,

  Please send out your talk abstracts in response to this!,. The Proposal deadline is 7/21 (Friday) 10AM IST.

Best Regards,

Santosh

   

[Aneesh Kumar]

On Mon, Jul 17, 2023 at 10:03 AM Santosh Shukla <santosh.s...@gmail.com> wrote:

Hi Bangalore Linux Kernel Community,

  Please send out your talk abstracts in response to this!,. The Proposal deadline is 7/21 (Friday) 10AM IST.

Is there a list of currently submitted proposals anywhere? I was expecting a reply to this email with talk abstracts. I don't see that yet. Is it collected by some other ways? 

 

-aneesh

[Ritesh Harjani]

Hi Aneesh, 

I submitted my abstract to Santosh, but didn't reply-all since there was an option for "reply to author". 

But here is my abstract for everyone else as well. 

Scaling filesystem block mapping using iomap and large folios 
=============================================================

Traditionally filesystem block mapping will first allocate & instantiate a page in the page cache, then for each page we will map a filesystem block.
This is suboptimal from a filesystem perspective because all modern filesystems are capable of allocating a single large extent which can span over multiple pages.

Hence iomap is the more modern way of mapping file offset range to extents that describe LBA range in block device address space. 
It is more filesystem centric abstraction rather than page cache centric and hence it can scale well. 
iomap will first map a single large extent and then for each such extent we instantiate folios in folio cache. 

In this talk I will talk about the pitfalls of older design and why should filesystems convert their I/O paths to use iomap.
Will share the conversion updates of ext2/ext4 & other filesystems to iomap and its current challenges.  
We will also discuss large folio support in iomap and our recent performance improvement done in iomap for improving database workloads
with large folio support or with large pagesize.

[1]: https://lore.kernel.org/linux-xfs/cover.1688188958...@gmail.com/
[2]: https://lore.kernel.org/linux-ext4/cover.1682069716...@gmail.com/

--
You received this message because you are subscribed to the Google Groups "Kernel Meetup Bangalore" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-meetup-ban...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/kernel-meetup-bangalore/CADGAMymcrwzkU3_crQQE8VL0vFJTYSViPodR4FOMAOHbt4hbUQ%40mail.gmail.com.

[Gautham]

Hello Santosh,

I would like to propose the following talk:

Queued Spinlocks in the Linux Kernel : Motivation, Design and Implementation

===========================================================

Linux Kernel has a elegant implementation of a variant of spin_lock named the queued_spinlock. 

This was motivated by the need to ensure fairness among the lock contenders, and also to ensure

scalability on platforms which support a large number of contenders.

In this talk we will explore the motivations, the design choices and the actual protocol with some examples.

Time : 40 mins. 

On Mon, Jul 17, 2023 at 10:03 AM Santosh Shukla <santosh.s...@gmail.com> wrote:

--

You received this message because you are subscribed to the Google Groups "Kernel Meetup Bangalore" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-meetup-ban...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/kernel-meetup-bangalore/8f8ca8fe-de77-4868-b8c3-40f16413c677n%40googlegroups.com.

--

Thanks and Regards
gautham.

[Bharata B Rao]

Here's my proposal:

Hardware hints for optimal page placement

 

Hardware platforms have started exposing useful and actionable memory access information to the OS in various ways.  There are sub-systems in the kernel like NUMA balancing which need such information but currently employ software methods to generate and act upon such access information. They could benefit if hardware can directly provide access information to the kernel in an easy to consume manner.

 

This talk will discuss the experience of using Instruction Based Sampling (IBS) mechanism present in AMD EPYC processors for NUMA balancing and hot page promotion.

Regards,

Bharata.

[Santosh Shukla]

Hi Aneesh,

 I am receiving talks/proposals and currently collating. I am going to share the shortlisted talk/proposal by tomorrow EOD.

Thanks,

Santosh

 

-aneesh

[Praveen Kumar]

If we have any slot, one of the topic which we would like to share and discuss is "Linux with Hyper-V (https://docs.kernel.org/virt/hyperv/overview.html). 

In this topic we would like to discuss about the high level architecture of Linux kernel stack on Hyper-V.

Regards,

~Praveen.

[Tarun Sahu]

Hi, Please find my abstract attached. +added the list Improve system performance by resolving internal fragmentation of Transparent Hugepages using hardware counters. ======================================================================================== Transparent hugepages are employed to enhance performance by decreasing TLB misses and LRU lock contention. However, internal fragmentation of THP can result in increased memory consumption, negatively impacting system performance. It is not feasible to determine the access count of individual subpages within a THP to assess internal fragmentation and decide whether the THP should be split and unused/cold subpages swapped out. Power platform features a distinct hardware counter that records access counts at the page level. While reference counts only indicate if a page has been recently accessed, hardware counters enable quantification of page access. This allows us to obtain per subpage access counts for a THP and measure its utilization, facilitating decisions on internal fragmentation of THP. We tested this with custom benchmark and memtier (with memcaced) benchmark and observed improvement in system performance with this feature enabled. In this talk, I will discuss the problem of internal fragmentation in Transparent Hugepages and share my findings on how it can be resolved using hardware counters, along with the results of my research on this topic. Regards, Tarun Sahu (IBM LTC)

[Sourabh jain]

Hello Satosh,

Here is the abstract of my presentation.

Title:

--------

Kdump hardening during CPU and memory hotplug events

Abstract:

---------------

Kdump is a crucial Linux kernel infrastructure that captures memory state

during kernel crashes, enabling effective system crash analysis. In a virtualized

environment, the hot-plugging of CPU and memory resources poses significant

challenges for kdump. This hinders its ability to collect valid dumps, consecutively

impacting serviceability.

The current solution to keep kdump updated with CPU/memory events is inefficient

and creates a considerable window during which kdump fails to collect the valid dump.

This presentation introduces an upstream proposed feature designed to improve the

kdump infrastructure's resilience during CPU and memory hotplug events, addressing

these limitations and ensuring more robust/reliable crash dump collection.

I believe I would require approximately 30 minutes in total to present this

topic and take few questions.

Thanks,

Sourabh Jain

--

[Ojaswin M]

Hi Santosh,

I'm interested in talking about the ext4 allocator work we've been doing recently. Please find the abstract below:

Filesystem block allocation techniques and improvements ======================================================= Filesystem allocator performance is generally optimized for all common types of workloads. Using various techniques like block preallocation / reservation, delayed allocations etc. filesystems try very hard to optimize block allocation while keeping fragmentation to a minimum. These techniques have been working well for several years now. But recently there is a growing demand of extreme use cases / data patterns. That is - 1. Large scale enterprise data centers require large filesystems with millions of block groups. 2. There is an increasing demand of small filesystem on a small disk space for cloud usecase due to pay-as-you-go model. (price per size of disk usage) This causes a lot of stress on filesystem to find a suitable block/extent for allocation request in order to keep fragmentation to minimum. This in turn wastes a lot of CPU cycles to find a suitable block and thus reduces the system performance. In this talk I would like to talk about various filesystem block allocation techniques and the recent multi-block allocator improvement work which we did for ext4. I will present various bottlenecks observed in different workload pattern and some graphs of performance improvement numbers.

Upstream lore link: https://lore.kernel.org/linux-ext4/cover.16854497...@linux.ibm.com/

Regards,

Ojaswin

[Dhruva gole]

Hello Santosh,

I have been working on Low Power modes and wakeup sources recently. I would like to share a few things I have learned about wake IRQ's in linux, here's the abstract and title:

Wake IRQs demystified - How Interrupts Wake a System

Wake IRQs play a crucial role in power management, allowing a system to enter low power states while still being able to wake up when needed. However, understanding how interrupts wake a system can be challenging, especially for those new to power management.

 

In this talk, we will delve into the basics of wake IRQs, explaining what they are, and how to enable them.

 

Additionally, we will discuss how wake IRQs fit into the broader context of power management, including how it can constraint a system from entering certain low power states. We will also explore the relationship between wake IRQs and device drivers, and how to ensure that devices are correctly configured to generate wake IRQs.

[Praveen Kumar]

Adding further details.

Linux Enlightenment for Hyper-V

=========================

Microsoft's Hyper-V hypervisor consists primarily of a bare-metal hypervisor plus a virtual machine management service running in the parent (root) partition (roughly equivalent to KVM and QEMU, for example). The Guest VMs run in the child partitions. The Linux kernel has been fully enlightened to run as Guest VM on Hyper-V. Further, there are developments, of having Linux kernel as Root partition on Hyper-V (https://lwn.net/Articles/844957/).
In the discussion, we would like to share the high level architecture design for Linux Guest (enlightened) + Hyper-V virtualization stack and dig on the concepts around the same.

The session could span through 30-45 mins.

Regards,

~Praveen.

Microsoft Linux Systems Group (LSG)

--

You received this message because you are subscribed to the Google Groups "Kernel Meetup Bangalore" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-meetup-ban...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/kernel-meetup-bangalore/5af3ad3f-81a7-479d-a7e6-b5db016f0fe7n%40googlegroups.com.

[Anuj gupta]

Hello Santosh,
I would like to propose the following talk.

io_uring passthrough: New I/O path in the Linux Kernel
=============================================
NVMe storage continues to evolve with new features and command sets, which do not fit well within the Linux abstraction layers and/or face adoption challenges.
The NVMe passthrough interface allows new device-features to be usable without having to build block-generic commands.
However existing passthrough interface has remain tied to synchronous ioctl, which is a blocker for performance-centric usage scenarios.
To overcome this challenge, io_uring_cmd was added as a modern equivalent of asynchronous ioctl and made its way into the 5.19 kernel.
io_uring coupled with nvme passthrough provides a scalable alternative along with better performance compared to existing block I/O path.
This talk will cover implementation details, usage and performance benchmark of io_uring passthrough against existing block I/O path.

Best Regards,

Anuj Gupta

[Allen Pais]

The submissions this time are insanely brilliant. Each submission showcases a unique perspective and 

Touches upon compelling topics that has left me inspired and in awe of the talent in Bangalore.

I would like to thank everyone for coming forward with such remarkable submissions.

Since I will not be part of the upcoming meetup, I would like to Nominate, Aneesh(IBM), Vinod(Linaro), Vaidyanathan(IBM)

To help select the talks.

This is exciting and Thank you again AMD and fellow hackers for making this possible.

- Allen 

--
You received this message because you are subscribed to the Google Groups "Kernel Meetup Bangalore" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kernel-meetup-ban...@googlegroups.com.

To view this discussion on the web, visit https://groups.google.com/d/msgid/kernel-meetup-bangalore/0fedc677-08c8-423f-85df-497fa9a0dd31n%40googlegroups.com.

[Santosh Shukla]

Hello Everyone,

We have reached a deadline for talks submission. As Allen alluded to in his note, we are pleasantly surprised at the number and technical spread of talks being proposed.

That said, since we are organizing this event at an external venue, there is both a space and attendee constraint due to which we have not been able to accommodate everyone desirous of attending the event. Hopefully, the Bangalore Linux kernel community will find ways to overcome this good-to-have challenge.

On the topic of talks, as announced earlier, we will have a panel discussion featuring prominent kernel contributors attending the event. The are 5 slots for 40 minute talks with sufficient time for switchover and breaks. Due to constraints mentioned above, we cannot accommodate more entries at this time.

As organizers of the community event, we would like to have an equitable distribution of topics across organizations, technologies and architectures. There are a few proposals that are useful to everyone irrespective of these considerations - Rust, workqueues, qspinlock, for instance. Given that all proposals cannot be accommodated at this instance of the meetup, maybe some of these can flow into the next meetup instance. To decide on which talks make the cut, Aneesh Kumar (IBM), Vinod Koul (Linaro) and Nikunj Dadhania (AMD) would be the voting members for unbiased decision making and based on current upstream presence.  (Viresh, Gautham, Neeraj, Ritesh and Bharata are active upstream members, but have topics proposed and so are not part of the decision making team).

It would be great to brainstorm on how to keep this community vibrant - volunteers anybody to host the next meetup? 

Another thought to accommodate the proposals is to have a slot for lightening talks - everyone gets 5 minutes to pitch a topic, with a longer talk in a subsequent meetup.

Regards,

Santosh

[Shrikanth Hegde]

Hi. 
Please find our abstract for the talk on Scheduler related problem found in Virtual Machines.

--------------------------------------------------------------------------------------------------------------------------------------------------
Hypervisor Preemptions - Paravirtualized Scheduler
--------------------------------------------------------------------------------------------------------------------------------------------------
When Linux is running above a Hypervisor, there are more chances of CPU overcommit, i.e., more vCPUs
compared to the allotted Physical CPUs. In a high Utilization scenario, switching between vCPUs results in 

preemption. This problem gets amplified when there multiple such overcommitted VMs which share the 

same physical CPUs.

What are the different ways to handle this?
1. Pin vCPUs to Physical CPUs.
- Not scalable.
- Linux Scheduler has less control and flexibility.
2. Hint based Scheduling aka Para Virtualized Scheduler
- Hypervisor provides the Hint to the Guest Scheduler about the real current physical resource it can
really use.
This has its own set of challenges w.r.t to topology, load balance, wakeups.

--------------------------------------------------------------------------------------------------------------------------------------------------

How can the Guest scheduler perform better when some Underlying CPUs are ceded?

--------------------------------------------------------------------------------------------------------------------------------------------------
(Can we hint if CPU is available or not.)
Most times Guest OS cannot know whether vCPUs are Ceded or it's been preempted by the hypervisor.
In case of preempted, it is generally better not to schedule that on vCPUs since it may take more time
for that vCPU to get a pCPU back and start executing.
If it only Ceded (and other SMT threads are active) then vCPU can be used for wakeups. There are
architecture hooks available in vcpu_is_preempted which can be used to provide such hints.

--------------------------------------------------------------------------------------------------------------------------------------------------

Asymmetric topologies in VMs.

--------------------------------------------------------------------------------------------------------------------------------------------------
Asymmetric topologies are becoming more and more common. Arm has big.LITTLE, DynamiQ, while
Intel has ITMT, Intel Hybrid Clusters, PowerPC had P7 to name a few.
These Asymmetric topologies lead to Asymmetric Sched domains. Current scheduler doesn't handle
Asymmetric domains, since it balances based on idle_cpus only.

--------------------------------------------------------------------------------------------------------------------------------------------------
How do we match Guest topology with Host topology?

--------------------------------------------------------------------------------------------------------------------------------------------------
OSPM discussions (led by Dario Faggioli) had some interesting thoughts on how to match Guest
topology with the Host topology to gain better performance in the guest.
On a task wakeup on an unpinned VM, Since the underlying CPU may have changed should we look for
wake-affine at all?

Regards
Srikar Dronamraju/Shrikanth Hegde