back
115 comments
Linux added a busy polling feature for high performance networking. Most Linux software does not use it, but software used in datacenters (e.g. by CDNs) that does use it makes the system very energy inefficient when things are not busy. This patch gives the the kernel the ability to turn that off when not busy to regain energy efficiency until things become busy again.

The article name is somewhat misleading, since it makes it sound like this would also apply to desktop workloads. The article says it is for datacenters and that is true, but it would have been better had the title ended with the words “in datacenters” to avoid confusion.

It’s a lot more nuanced than that. Being in a data center doesn’t imply heavy network utilization. The caveat is clearly outlined in the article being workload based not deployment based. If you have a home machine doing network routing it would absolutely benefit from this. In fact I would say probably the vast majority of Linux installs are home network devices, just people don’t know it. Embedded Linux machines doing network routing or switching or IPS or NAS or whatever would benefit a lot from this. “Energy savings” can be seen as a green washing statement but on embedded budgets it’s a primary requirement.
> The article name is somewhat misleading, since it makes it sound like this would also apply to desktop workloads.

The article name is, Data Centers Can Slash Power Needs With One Coding Tweak: Reworking 30 lines of Linux code could cut power use by up to 30 percent

The article says, “It is sort of a best case because the 30 percent applies to the network stack or communication part of it,” Karsten explains. “If an application primarily does that, then it will see 30 percent improvement. If the application does a lot of other things and only occasionally uses the network, then the 30 percent will shrink to a smaller value.”

It seems you only read the HN title? If so, why bother to critique the article's title?

heh, will this become the equivalent of "in mice" for bio papers?
I appears that this academic is very good at public relations.

They mix interrupts and polling depending on the load. The interrupt service routine and user-kernel context-switch overhead is tiny computationally and hence in power usage.

Also, most network hardware in the last twenty years has had buffer coalescing, reducing interrupt rates.

But linux use outside datacenters and server farms is irrelevant in terms of power use anyways I don't think anybody would be confused in the context of power usage.
So it's not even on by default, it needs to be turned on explicitly?
you don’t have to say “in datacenters” when talking about linux, that is the obvious context in the vast majority of cases
One thing I didn’t see mentioned here yet: a lot of high-performance data center workloads don’t actually go through the Linux kernel’s network stack at all.

Instead, they use DPDK, XDP, or userspace stacks like Onload or VMA—often with SmartNICs doing hardware offload. In those cases, this patch wouldn’t apply, since packet processing happens entirely outside the kernel.

That doesn’t mean the patch isn’t valuable—it clearly helps in setups where the kernel is in the datapath (e.g., CDNs, ingress nodes, VMs, embedded Linux systems). But it probably won’t move the needle for workloads that already bypass the kernel for performance or latency reasons. So the 30% power reduction headline is likely very context-dependent.

There are a whole lot of commodity 1-2U rack mount boxes running RHEL or CentOS or the like out there that are mostly idle, and which don’t do anything fancier in hardware than maybe checksum or VLAN offload.
For a more detailed look at this change: https://lwn.net/Articles/1008399/
And for even more details, you can look at the relevant patch: https://lwn.net/ml/all/20241021015311.95468-1-jdamato@fastly...

The power efficiency seems to be limited to "network applications using epoll".

The 30% the article talks about seems to be benchmarked on memcached, and here is the ~30 lines diff they're probably talking about: https://raw.githubusercontent.com/martinkarsten/irqsuspend/m...

This is really cool. As a high-performance computing professional, I've often wondered how much energy is wasted due to inefficient code and how much that is a problem as planetary compute scales up.

For me, it feels like a moral imperative to make my code as efficient as possible, especially when a job will take months to run on hundreds of CPU.

Year ago, I posted here that there should be some sort of ongoing Green X-Prize for this style of Linux kernel optimization. It's still crazy to me that this doesn't exist.
Meanwhile I'm running four Electron apps just to act like different shitty IRCs.
> I've often wondered how much energy is wasted due to inefficient code and how much that is a problem as planetary compute scales up.

I personally believe the majority is wasted. Any code that runs in an interpreted language, JIT/AOT or not, is at a significant disadvantage. On performance measurements it's as bad as 2x to 60x worse than the performance of the equivalent optimized compiled code.

> it feels like a moral imperative to make my code as efficient as possible

Although we're still talking about fractions of a Watt of power here.

> especially when a job will take months to run on hundreds of CPU.

To the extent that I would say _only_ in these cases are the optimizations even worth considering.

Absolutely.

It is unfortunate that many software engineers continue to dismiss this as "premature optimization".

But as soon as I see resources or server costs gradually rising every month (even on idle usage) costing into the tens of thousands which is a common occurrence as the system scale, then it becomes unacceptable to ignore.

My experience with HPC is only tangential to being a sysadmin for data taking and cluster management for a high energy physics project; I am interested on your thoughts about using generative AI to search out for potentially power inefficient code paths in codebases for potential improvement.
Does this mean that "adaptive interrupt mitigation" is no longer a thing in the kernel? I haven't really messed with it in ~15+ years, but it used to be that the kernel would adapt, if network rate was low it would use interrupts, but then above a certain point it would switch to turning off interrupts and using polling instead.

The issue I was trying to resolve was sudden, dramatic changes in traffic. Think: a loop being introduced in the switching, and the associated packet storm. In that case, interrupts could start coming in so fast that the system couldn't get enough non-interrupted time to disable the interrupts, UNLESS you have more CPUs than busy networking interfaces. So my solution then was to make sure that the Linux routers had more cores than network interfaces.

Adaptive IRQ moderation is a hardware feature. https://edc.intel.com/content/www/us/en/design/products/ethe...
The flip side of this is meta having a hack that keeps their GPUs busy so that the power draw is more stable during llm training (eg don't want a huge power drop when synchronizing batches)
I thought of this too. Iirc it was bigger problem because surging spikes in power and cooling were harder and more costly to account for.

I'm not au fait with network data centres though, how similar are they in terms of their demands?

Have a link to an article talking about them doing this?
That goes against separation of concerns. A separate utility must be created for that specific purpose, not hidden in some other part of the system
Off topic: glad to read about Joe Damato again — such a blast from the past. I haven't read anything from him since I first read James Gollick posts about on tcmalloc and then learning about packagecloud.io which eventually led me to Joe's amazing posts.
Key paragraph: "This energy savings comes with a caveat. “It is sort of a best case because the 30 percent applies to the network stack or communication part of it,” Karsten explains. “If an application primarily does that, then it will see 30 percent improvement. If the application does a lot of other things and only occasionally uses the network, then the 30 percent will shrink to a smaller value.”"
I'd suggest a submission title change to "...cut networking software power use..."

Not so sexy. But as a force multiplier that's still a lot of carbon probably.

When your sentence contains "Up To", literally anything is possible.
This brings back memories.

https://didgets.substack.com/p/finding-and-fixing-a-billion-...

Wow busy waiting is more resource intensive than I realized.
When a machine is large enough with enough tenants, there should be enough ambient junk going on to poll on 1 or a few cores and still be optimized for power. This is the tradeoff described in the 2019 "SNAP" paper from Google, in the section about the compacting engine.

The "up to 30%" figure is operative when you have a near-idle application that's busy polling, which is already dumb. There are several ways to save energy in that case.

> The "up to 30%" figure is operative when you have a near-idle application that's busy polling, which is already dumb.

That was my first thought, but it sounds like the OS kernel, not the application, has control over the polling behavior, right?

I thought that Intel added the ‘pause’ instruction to make busy spinning more power friendly
x86 has always had PAUSE, what they did is they made the PAUSE way longer in Skylake-X which threw everyone off guard. But yeah, x86 ISA extensions cover this ground very well. Atom, the low-power server line, introduced UMWAIT and TPAUSE that enable cores to briefly enter a power-saving active state while waiting for something to happen. These instructions later came to mainstream Core and Xeon CPUs because Intel made those frankenchips with Atoms and Cores together.
> When I grew up in computer science in the 90s, everybody was concerned about efficiency...“Somehow in the last 20 years, this has gotten lost. Everybody’s completely gung-ho about performance without any regard for efficiency or resource consumption. I think it’s time to pivot, where we can find ways to make things a little more efficient.

This is the sort of performance efficiencies I want to keep seeing on this site, from those who are distinguished experts and contributed to critical systems such as the Linux kernel.

Unfortunately, in the last 10-15 years we are seeing the worst technologies being paraded due to a cargo-cultish behaviour. From asking candidates to implement the most efficient solution to a problem in interviews but then also choosing the most extremely inefficient technologies to solve certain problems because so-called software shops are racing for that VC money. Money that goes to hundreds of k8s instances on many over-provisioned servers instead of a few.

Performance efficiency critically matters, and it is the difference between having enough runway for a sustainable business vs having none at all.

And nope. AI Agents / Vibe coders could not have come up with a more correct solution in the article.

I love how I can update Linux and expect it to get faster, not slower. Exception to be made for Nvidia drivers, but at least their performance is consistent.
Oh, as a guy which was using DPDK like technology that do busy-poll and bypass kernel to process network packets, I must say may be much power have be wasted...
Standard practice in all trading applications is to busy poll the NIC using a kernel bypass technology.

Typically saves 2-3 microseconds going through the kernel network stack.

Noob question: how do you enable this power saving feature? Is it just upgrading to the latest kernel?
I'd like to see clearly explained Linux boot code that does not invoke a prayer to Satan.
Niiice. What if we reworked 100 lines?
A manager went to the master programmer and showed him the requirements document for a new application. The manager asked the master: “How long will it take to design this system if I assign five programmers to it?”

“It will take one year,” said the master promptly.

“But we need this system immediately or even sooner! How long will it take if I assign ten programmers to it?”

The master programmer frowned. “In that case, it will take two years.”

“And what if I assign a hundred programmers to it?”

The master programmer shrugged. “Then the design will never be completed,” he said.

— Chapter 3.4 of The Tao of Programming, by Geoffrey James (1987)

20%.
Leverage porn
> The code has been implemented as part of the Linux kernel release version 6.13 in January.

This is basically 3 month old news now.