A natural r&d project would be a high frequency oms or something like an nyse bid/ask/match book manager.
These apps tend to do kernel bypass anyway for net i/o after hard scrubbing the os to remove as many interrupts and other superfluous jitter the oms does not need.
However, I believe the low level kernel bypass code (eg dpdk, libfabric) depend on Linux to talk to the hw in a disciplined way.
Anybody know better?
A decently faster oms this way could be a practical alternative to fpgas we sometimes see in that space. (hw would be co-located of course)
If you have serious attack surface minimisation goals you need to embrace FOGAs. LLM coded unikernels have way too much bloat.
The Zynq UltraScale+ parts pair a hard ARM CPU with FPGA fabric. You can progressively isolate fast path and high security components in logic. LLMs are getting quite good at using synthesis tools.
I've had limited adventures in embedded software and as part of that became a fan of Zephyr (http://www.zephyrproject.org) and wonder about its potential application as a unikernel.
On the minus side: I think porting "traditional" software - say Nginx - to it will be agonising. And I can't make any guarantees about its performance in the same way that Alpine Linux container images are not necessarily fast.
But there's an entire toolchain, debugging story, drivers etc. for something that compiles to tiny and starts up basically instantaneously. Surely that's got to be useful for something?
Reducing attack surface is definitely a plus but it is nowhere close to the number one security benefit of running unikernels.
That's why I never really liked talking about "reducing attack surface" that much because folk inevitably turn to lines and code, which while reducing is good, just simply doesn't communicate what the biggest problem truly is.
Vuln exploitation is the number one entry point for data breaches and os command injection is the number one CWE in CISA Kev from last year.
System intrusion was repeated something like 64 times in last year's DBIR.
The operating system itself is literally the problem as it's inherently meant to run many different programs whereas unikernels only run one.
this question always comes up. I personally think the status quo is pretty weak here. I have certainly added a gdb stub to a unikernel before, but that apparently given that it was never used it wasn't the right answer.
people always talk about losing the excellent debugging facilities that linux gives you. if a service crashes in a big cloud environment what is that you do now that works so well. do you think its intractable to implement that in a single-process kernel?
I don't know if anything like this has been built, but I'd imagine you could have the hypervisor capture stack traces / log buffers / core dumps when a unikernel VM crashes.
I'm not a security researcher but I know that VMs have been the tool of choice to step through malware execution for at least the past 15 years. I recall coming across ida extensions for it.
Yeah that's generally the argument for using a managed runtime (like Ocaml in MirageOS's case, or I could see Go or even the JVM fitting here) for these kinds of things. Running a VM which has no pointers, memory access etc primitives, and is garbage collected etc direct on "metal" gives more peace of mind about that sort of thing.
You could make the argument that Rust w/ its memory safety is a candidate, but w/ Rust it's still entirely possible and fairly easy to break out of that. And Rust/Cargo applications have a habit of using a bazillion third party deps that you then need to keep a close eye on.
For that matter, the actual Linux kernel need not be large at all. It's really the user space stuff that takes up the bulk of what people think of as the largess of Linux.
Still, there's a whole pile of stuff there that many applications don't need. If you truly can run the entire of the OCaml runtime and standard libraries etc without it, and you're happy writing your apps in OCaml... MirageOS has always felt winning-ish to me as a general idea.
And to TFA's point, now with LLMs the relative exotic nature of OCaml as a platform need not be a huge hindrance. I like the language, but have only dipped my toes in. Definitely tempting to try writing apps on it w/ MirageOS now. Though I doubt I'd convince anybody to pay me to do that...
I was thinking that we don't need programs to share CPUs anymore now that each machine has so many of them. We could have an OS which binds each program to one CPU. When it runs out of CPUs, it suggests to close an existing program. Then programs don't need to pay context-switching cost.
I’m completely ignorant about this and likely just missing the point but: isn’t the point of an OS to not have to (vibe)code filesystems and networking by hand every time you need them? Also, how does a unikernel cooperate with other applications? Would they all live in separate networked unikernels managed by a hypervisor? And, if they all (vibe)coded their own fs and network wouldn’t that introduce subtle bugs and inconsistencies that would eventually bring the whole thing down and lead you back to the need for shared primitives in the first place?
Maybe a good halfway point is to still have a unikernel but the libraries for common stuff like networking or fs are already written according to a standard and plug and play and reusable across applications?
Unikernels simply redraw abstraction boundaries. You can still use libraries and third-party code to implement common primitives. In fact, that's exactly what many unikernel frameworks give you (unikraft, hermit, mirage)
I guess the others have made these points, but since unikernels run in VMs with virtual hardware, I guess a virtual ethernet card can have an interface that's easy to program for, making the driver trivial. You're not talking to real hardware in any case.
Similarly, a filesystem's just a data structure that happens to be on disc. You can mmap your harddrive and let the host hypervisor handle swap, preemption etc.
It's similar to a process contract, but with slightly different boundaries
The thing I don't get: We'll now run those N unikernel things in VMs with a hypervisor underneath. How is that different from running static binaries in a locked-down OS, i.e., with a traditional kernel? You can surely break out of a unikernel by "popping the kvm" as the article mentions. How is that more secure? In principle we are running the same things, just call them differently (app becomes unikernel+app, kernel becomes hypervisor).
You don't build the networking or filesystem by hand every time you need them. You pull them from reviewed and maintained library code from a repository. That's how MirageOS etc work. Someone else has written e.g. TCP/IP -- hopefully well -- and you link against it.
The point that makes this different from an OS is that there's no shared service, no syscalls to do that stuff -- just subroutine calls -- and no timesharing (except at the hypervisor level). It's just a single runtime running on hypervisor direct against the virtualized hardware.
Because, yeah, all this work has happened in hardware in the last 30 years to make that possible, but we still treat the operating system as the best unit of resource sharing. When in fact it's kind of a jack of all trades master of none. If you're booting a whole linux kernel -- with its giant framework built to co-tenant a pile of applications and users -- just to run one process, there's definitely something to look at there. Even if the unikernel space itself is relatively immature.
You don't need to emphasize the "vibe coded every time you need them" thing. That's not what anybody was being talked about in the article. The "vibe coding" piece is pointing out that the missing pieces in the ecosystem can be more rapidly filled in now because of agentic tooling.
Over time, what prevents the reviewed-and-maintained-library-code from developing creeping featuritis and consequently evolving the additional attack surface that unikernels currently lack?
> running on hypervisor direct against the virtualized hardware.
Well, here is your actual OS then. You named it "hypervisor", and changed the system interface, but it's still there, revirtualizing the access to hardware and ensuring separation of different users/applications/unikernels.
This isn't too dissimilar to DOS programs. DOS also gave you a filesystem, memory management, a shell a process loder, driver interface, but it was all in real mode, and you were free to replace these components with your own as you pleased.
Or if you're more familiar with embedded, like a BSP or RTOS
Arbitrating access to the hardware between multiple tenants still gonna be a PITA (as it was during the DOS days), especially if you want to expose actual hardware, not some generic VirtIO devices. Even with the latter, the hypervisor would have to do quite a bit of the resource management and interface translation.
Unless you're literally running a single application on a dedicated piece of hardware in which case... yeah, sure, you can do the OS-as-a-library schtick. But you could do it 10 years ago, too.
Does it though? And I'd really like to see how you'll revirtualize e.g. an NVIDIA GPU (notoriously context-full piece of hardware), while allowing transparent shared access to it from different applications. Heck, you'll run into trouble with most of PCI-connected devices except for the most primitive once.
yea, "virtualizing" an NVIDIA GPU is not truly feasible right now from what I see.
I do know that AMD has done more in this space. When I worked at Google there were (other) people on the Stadia team doing this. And some open source bits out there that I've seen, including from AMD.
Unfortunately, my real world work... and most real inference etc work others are doing... is on CUDA/NVIDIA.
Intuitively, I think I have some appreciation for how much silicon has been poured into virtualization. Its always seemed like a "yeah but you could avoid those costs by not doing that" but I'm more receptive to the idea that these might in some cases be really good ways to get some of the isolation workloads demand with the hardware helping us out, these days. There's so much securing for vm's, and maybe it's just easier than trying to secure in userlands: let the hardware help.
Long time interest in microvm's but they felt heavier weight than I wanted. Now it feels like maybe they might actually in some regards in some ways be lighter weight than managing workloads yourself in userland. Maybe. I dunno. Interesting times, i'm open to it.
One of the problems is there's only a small certain percentage of scenarios/applications that benefit from very short start times.
I'd wager most of the services running on the interwebs are web servers, database servers, inference servers etc that don't change over their lifetime really. They're just doing the same thing all day long every day.
If you're doing stuff like what I'm doing for work right now, which is, yeah, multiplexing potentially oodles of user-submitted jobs, and those jobs are best expressed as distinct images or containers, then yes, managing start times is absolutely imperative in improving utilization/occupancy and therefore reducing costs.
But I'm not convinced that's a typical scenario, not typical enough to drive enough time and money investment in this space maybe?
Also there ain't currently no real "hypervisor" for the (NVIDIA) GPU. Not practically anyways. And that's arguably where we need it the most. Or at least I do, for Day Job(tm).
So unikernels and microvms may have to lean on other arguments for adoption: security and simplicity-to-reason-about might be those...
A natural r&d project would be a high frequency oms or something like an nyse bid/ask/match book manager.
These apps tend to do kernel bypass anyway for net i/o after hard scrubbing the os to remove as many interrupts and other superfluous jitter the oms does not need.
However, I believe the low level kernel bypass code (eg dpdk, libfabric) depend on Linux to talk to the hw in a disciplined way.
Anybody know better?
A decently faster oms this way could be a practical alternative to fpgas we sometimes see in that space. (hw would be co-located of course)
If you have serious attack surface minimisation goals you need to embrace FOGAs. LLM coded unikernels have way too much bloat.
The Zynq UltraScale+ parts pair a hard ARM CPU with FPGA fabric. You can progressively isolate fast path and high security components in logic. LLMs are getting quite good at using synthesis tools.
https://www.amd.com/en/products/system-on-modules/kria/k26/k...
It’s good to see this! I’m offering a cloud service to host unikernels based on my BareMetal kernel.
~$0.005 CAD/h for a small network-enabled VM with 4MiB of RAM and no disk.
https://baremetal.returninfinity.com
I've had limited adventures in embedded software and as part of that became a fan of Zephyr (http://www.zephyrproject.org) and wonder about its potential application as a unikernel.
On the plus side: it's a unikernel. You compile it and your application together and get a binary. It has support for running on virtual hardware (https://docs.zephyrproject.org/latest/hardware/virtualizatio...) and virtio is the new bios, right?
On the minus side: I think porting "traditional" software - say Nginx - to it will be agonising. And I can't make any guarantees about its performance in the same way that Alpine Linux container images are not necessarily fast.
But there's an entire toolchain, debugging story, drivers etc. for something that compiles to tiny and starts up basically instantaneously. Surely that's got to be useful for something?
Reducing attack surface is definitely a plus but it is nowhere close to the number one security benefit of running unikernels.
That's why I never really liked talking about "reducing attack surface" that much because folk inevitably turn to lines and code, which while reducing is good, just simply doesn't communicate what the biggest problem truly is.
Vuln exploitation is the number one entry point for data breaches and os command injection is the number one CWE in CISA Kev from last year.
System intrusion was repeated something like 64 times in last year's DBIR.
The operating system itself is literally the problem as it's inherently meant to run many different programs whereas unikernels only run one.
> The operating system itself is literally the problem as it's inherently meant to run many different programs whereas unikernels only run one.
Unless you rely on security through compartmentalization. See: https://qubes-os.org
As mentioned before on this topic. What about debug ability? An application overflow now corrupts part of the network stack.
In an oxide episode there were some mentions of reading off data lines but I just don’t think that’s practical.
The reduced attack service is cool but not at the expense of my visibility and liveness of the system
this question always comes up. I personally think the status quo is pretty weak here. I have certainly added a gdb stub to a unikernel before, but that apparently given that it was never used it wasn't the right answer.
people always talk about losing the excellent debugging facilities that linux gives you. if a service crashes in a big cloud environment what is that you do now that works so well. do you think its intractable to implement that in a single-process kernel?
I don't know if anything like this has been built, but I'd imagine you could have the hypervisor capture stack traces / log buffers / core dumps when a unikernel VM crashes.
I'm not a security researcher but I know that VMs have been the tool of choice to step through malware execution for at least the past 15 years. I recall coming across ida extensions for it.
Yeah that's generally the argument for using a managed runtime (like Ocaml in MirageOS's case, or I could see Go or even the JVM fitting here) for these kinds of things. Running a VM which has no pointers, memory access etc primitives, and is garbage collected etc direct on "metal" gives more peace of mind about that sort of thing.
You could make the argument that Rust w/ its memory safety is a candidate, but w/ Rust it's still entirely possible and fairly easy to break out of that. And Rust/Cargo applications have a habit of using a bazillion third party deps that you then need to keep a close eye on.
Unikernels are on the smaller side of the spectrum; Linux on the larger side, with a lot of room in the middle...
For that matter, the actual Linux kernel need not be large at all. It's really the user space stuff that takes up the bulk of what people think of as the largess of Linux.
Still, there's a whole pile of stuff there that many applications don't need. If you truly can run the entire of the OCaml runtime and standard libraries etc without it, and you're happy writing your apps in OCaml... MirageOS has always felt winning-ish to me as a general idea.
And to TFA's point, now with LLMs the relative exotic nature of OCaml as a platform need not be a huge hindrance. I like the language, but have only dipped my toes in. Definitely tempting to try writing apps on it w/ MirageOS now. Though I doubt I'd convince anybody to pay me to do that...
I was thinking that we don't need programs to share CPUs anymore now that each machine has so many of them. We could have an OS which binds each program to one CPU. When it runs out of CPUs, it suggests to close an existing program. Then programs don't need to pay context-switching cost.
I’m completely ignorant about this and likely just missing the point but: isn’t the point of an OS to not have to (vibe)code filesystems and networking by hand every time you need them? Also, how does a unikernel cooperate with other applications? Would they all live in separate networked unikernels managed by a hypervisor? And, if they all (vibe)coded their own fs and network wouldn’t that introduce subtle bugs and inconsistencies that would eventually bring the whole thing down and lead you back to the need for shared primitives in the first place?
Maybe a good halfway point is to still have a unikernel but the libraries for common stuff like networking or fs are already written according to a standard and plug and play and reusable across applications?
Unikernels simply redraw abstraction boundaries. You can still use libraries and third-party code to implement common primitives. In fact, that's exactly what many unikernel frameworks give you (unikraft, hermit, mirage)
I guess the others have made these points, but since unikernels run in VMs with virtual hardware, I guess a virtual ethernet card can have an interface that's easy to program for, making the driver trivial. You're not talking to real hardware in any case.
Similarly, a filesystem's just a data structure that happens to be on disc. You can mmap your harddrive and let the host hypervisor handle swap, preemption etc.
It's similar to a process contract, but with slightly different boundaries
The thing I don't get: We'll now run those N unikernel things in VMs with a hypervisor underneath. How is that different from running static binaries in a locked-down OS, i.e., with a traditional kernel? You can surely break out of a unikernel by "popping the kvm" as the article mentions. How is that more secure? In principle we are running the same things, just call them differently (app becomes unikernel+app, kernel becomes hypervisor).
You don't build the networking or filesystem by hand every time you need them. You pull them from reviewed and maintained library code from a repository. That's how MirageOS etc work. Someone else has written e.g. TCP/IP -- hopefully well -- and you link against it.
The point that makes this different from an OS is that there's no shared service, no syscalls to do that stuff -- just subroutine calls -- and no timesharing (except at the hypervisor level). It's just a single runtime running on hypervisor direct against the virtualized hardware.
Because, yeah, all this work has happened in hardware in the last 30 years to make that possible, but we still treat the operating system as the best unit of resource sharing. When in fact it's kind of a jack of all trades master of none. If you're booting a whole linux kernel -- with its giant framework built to co-tenant a pile of applications and users -- just to run one process, there's definitely something to look at there. Even if the unikernel space itself is relatively immature.
You don't need to emphasize the "vibe coded every time you need them" thing. That's not what anybody was being talked about in the article. The "vibe coding" piece is pointing out that the missing pieces in the ecosystem can be more rapidly filled in now because of agentic tooling.
Over time, what prevents the reviewed-and-maintained-library-code from developing creeping featuritis and consequently evolving the additional attack surface that unikernels currently lack?
> running on hypervisor direct against the virtualized hardware.
Well, here is your actual OS then. You named it "hypervisor", and changed the system interface, but it's still there, revirtualizing the access to hardware and ensuring separation of different users/applications/unikernels.
This isn't too dissimilar to DOS programs. DOS also gave you a filesystem, memory management, a shell a process loder, driver interface, but it was all in real mode, and you were free to replace these components with your own as you pleased.
Or if you're more familiar with embedded, like a BSP or RTOS
Arbitrating access to the hardware between multiple tenants still gonna be a PITA (as it was during the DOS days), especially if you want to expose actual hardware, not some generic VirtIO devices. Even with the latter, the hypervisor would have to do quite a bit of the resource management and interface translation.
Unless you're literally running a single application on a dedicated piece of hardware in which case... yeah, sure, you can do the OS-as-a-library schtick. But you could do it 10 years ago, too.
I mean, sure we can put labels wherever we want. I don't really care. MirageOS even has "OS" in its name.
The point is all the other things: total number of lines of code, permissions and security, memory management, drivers, process mgmt, it all changes.
Does it though? And I'd really like to see how you'll revirtualize e.g. an NVIDIA GPU (notoriously context-full piece of hardware), while allowing transparent shared access to it from different applications. Heck, you'll run into trouble with most of PCI-connected devices except for the most primitive once.
yea, "virtualizing" an NVIDIA GPU is not truly feasible right now from what I see.
I do know that AMD has done more in this space. When I worked at Google there were (other) people on the Stadia team doing this. And some open source bits out there that I've seen, including from AMD.
Unfortunately, my real world work... and most real inference etc work others are doing... is on CUDA/NVIDIA.
I do wonder what kind of wins we're going to see from unikernels.
I used to regard V8 Isolates as a best possible sort of technology, with userlands juggling lots of processes.
Seeing netlify & unikraft switch to microvm's and have such a huge speed up was a bit of an awakening for me. Those are really fast start times! https://www.netlify.com/blog/edge-functions-firecracker-micr... https://unikraft.com/customer-stories/edge-functions-netlify...
Intuitively, I think I have some appreciation for how much silicon has been poured into virtualization. Its always seemed like a "yeah but you could avoid those costs by not doing that" but I'm more receptive to the idea that these might in some cases be really good ways to get some of the isolation workloads demand with the hardware helping us out, these days. There's so much securing for vm's, and maybe it's just easier than trying to secure in userlands: let the hardware help.
Long time interest in microvm's but they felt heavier weight than I wanted. Now it feels like maybe they might actually in some regards in some ways be lighter weight than managing workloads yourself in userland. Maybe. I dunno. Interesting times, i'm open to it.
Edit: I just chatted with Astra Pro some about these ideas, if anyone wants some sense material to chew on. https://chatgpt.com/share/6aca924c-8e64-83ea-a5c3-aae08b2cdf...
One of the problems is there's only a small certain percentage of scenarios/applications that benefit from very short start times.
I'd wager most of the services running on the interwebs are web servers, database servers, inference servers etc that don't change over their lifetime really. They're just doing the same thing all day long every day.
If you're doing stuff like what I'm doing for work right now, which is, yeah, multiplexing potentially oodles of user-submitted jobs, and those jobs are best expressed as distinct images or containers, then yes, managing start times is absolutely imperative in improving utilization/occupancy and therefore reducing costs.
But I'm not convinced that's a typical scenario, not typical enough to drive enough time and money investment in this space maybe?
Also there ain't currently no real "hypervisor" for the (NVIDIA) GPU. Not practically anyways. And that's arguably where we need it the most. Or at least I do, for Day Job(tm).
So unikernels and microvms may have to lean on other arguments for adoption: security and simplicity-to-reason-about might be those...