The State of ARM RISC in the DataCenter
-
@Dashrender said in What Are You Doing Right Now:
@scottalanmiller said in What Are You Doing Right Now:
@Dashrender said in What Are You Doing Right Now:
Sure, I see when we don't have typical 2-4 socket computers, instead a server might have 20+ sockets.
Actually the move will be to single sockets, at least at first. ARMs rarely support multiple sockets.
So how do you see these appearing in the DC? Single socket, I'm guessing you're not virtualizing.
Why? It can go both ways, but Xen and containers are the standards that are expected. Why would single socket not have you virtualizing?
-
@Dashrender said in What Are You Doing Right Now:
@scottalanmiller said in What Are You Doing Right Now:
@MattSpeller said in What Are You Doing Right Now:
@thwr @Dashrender It will be interesting to see how it works out between consolidation (ARM server racks) and IoT/shards/stand-alone/single-board-computers
SBCs are the expectation for racks of ARM servers, like MoonShot.
This makes great sense when looking at DevOps, but I don't understand how they would work in a typical virtualized setup - but maybe that's not who they are going up against?
DevOps and typical virtualization overlap. These are smaller individual units that you are used to, but why do you feel that this would significantly impact virtualization or deployment decisions outside of needing more nodes that are cheaper rather than fewer that are more expensive?
-
@Dashrender said in What Are You Doing Right Now:
@scottalanmiller said in What Are You Doing Right Now:
@MattSpeller said in What Are You Doing Right Now:
@thwr @Dashrender It will be interesting to see how it works out between consolidation (ARM server racks) and IoT/shards/stand-alone/single-board-computers
SBCs are the expectation for racks of ARM servers, like MoonShot.
This makes great sense when looking at DevOps, but I don't understand how they would work in a typical virtualized setup - but maybe that's not who they are going up against?
Don't forget that ARM SoCs can easily pack dozens of cores in a single package. For example: https://www.nextplatform.com/2016/09/01/details-emerge-chinas-64-core-arm-chip/
I would expect to see servers like the aforementioned MoonShot from HP: Lots of cores per SoC, but just one SoC per board. And dozens of boards per chassis.
-
@thwr said in What Are You Doing Right Now:
@Dashrender said in What Are You Doing Right Now:
@scottalanmiller said in What Are You Doing Right Now:
@MattSpeller said in What Are You Doing Right Now:
@thwr @Dashrender It will be interesting to see how it works out between consolidation (ARM server racks) and IoT/shards/stand-alone/single-board-computers
SBCs are the expectation for racks of ARM servers, like MoonShot.
This makes great sense when looking at DevOps, but I don't understand how they would work in a typical virtualized setup - but maybe that's not who they are going up against?
Don't forget that ARM SoCs can easily pack dozens of cores in a single package. For example: https://www.nextplatform.com/2016/09/01/details-emerge-chinas-64-core-arm-chip/
I would expect to see servers like the aforementioned MoonShot from HP: Lots of cores per SoC, but just one SoC per board. And dozens of boards per chassis.
Exactly. Blade approach, but very dense. So maybe 16 servers in a chassis. Each server might be 64 cores. So a 3U user might be really dense in cores.
-
@scottalanmiller said in What Are You Doing Right Now:
@Dashrender said in What Are You Doing Right Now:
@scottalanmiller said in What Are You Doing Right Now:
@MattSpeller said in What Are You Doing Right Now:
@thwr @Dashrender It will be interesting to see how it works out between consolidation (ARM server racks) and IoT/shards/stand-alone/single-board-computers
SBCs are the expectation for racks of ARM servers, like MoonShot.
This makes great sense when looking at DevOps, but I don't understand how they would work in a typical virtualized setup - but maybe that's not who they are going up against?
DevOps and typical virtualization overlap. These are smaller individual units that you are used to, but why do you feel that this would significantly impact virtualization or deployment decisions outside of needing more nodes that are cheaper rather than fewer that are more expensive?
Mostly because of the added overhead of a hypervisor for every processor. If you look at modern server today, has two sockets and 8+ cores. So that's like having 16+ processors.
I know ARM chips can have multiple cores, so I'm not sure where the problem is running multisocket is compared to multicore. But I did infer (probably wrongly) that a single ARM would be single core, and now you're talking about 16x the overhead for 16 versions of XS running to support all of those processors.Also, How do you give a single process more power over, for lack of a better term, separate machines?
I'm talking completely out of my ass because I don't understand the architecture of ARM systems compared to Intel.
-
@Dashrender said in What Are You Doing Right Now:
Mostly because of the added overhead of a hypervisor for every processor. If you look at modern server today, has two sockets and 8+ cores. So that's like having 16+ processors.
Sort of. There are a few things that you are missing here....
- Additional sockets create additional overhead that is different than cores.
- Even enterprise AMD64 servers today have a trend towards single sockets (Scale is a big leader in that direction)
- In the big iron RISC world, single socket Power and Sparc systems are extremely common.
- ARM RISC can be 64 cores (real cores) on a single CPU, more than on AMD64
- Hypervisor overhead is absolutely tiny and because these are servers, Xen and PV are the norm with nearly zero overhead
- Containers have even lower overhead than PV
-
@Dashrender said in What Are You Doing Right Now:
I know ARM chips can have multiple cores, so I'm not sure where the problem is running multisocket is compared to multicore.
Multisocket and multicore have very different physical engineering needs. This is why standard cheap AMD64 procs (like the ones that you get in a desktop) can only be one socket. They don't have the channels for the dual sockets. And then the next step up Intel and AMD AMD64 processors are dual socket capable. These are the ones in most servers. They are cheaper than the ones that can do four and eight way sockets. Each growth in socket count makes the processors more expensive. This is why four and eight way AMD64 servers are not that common, the processor costs go up a lot and the efficiency goes down.
Multicore is almost entirely a software problem to solve. Multisocket is both software and hardware together. It has dramatically more overhead than just adding cores.
-
@scottalanmiller said in What Are You Doing Right Now:
@Dashrender said in What Are You Doing Right Now:
Mostly because of the added overhead of a hypervisor for every processor. If you look at modern server today, has two sockets and 8+ cores. So that's like having 16+ processors.
Sort of. There are a few things that you are missing here....
- Additional sockets create additional overhead that is different than cores.
- Even enterprise AMD64 servers today have a trend towards single sockets (Scale is a big leader in that direction)
- In the big iron RISC world, single socket Power and Sparc systems are extremely common.
- ARM RISC can be 64 cores (real cores) on a single CPU, more than on AMD64
- Hypervisor overhead is absolutely tiny and because these are servers, Xen and PV are the norm with nearly zero overhead
- Containers have even lower overhead than PV
Should we fork that whole topic? Contains quite a bit of information about ARM based servers and ARM vs Intel now. The discussion started with ~ https://mangolassi.it/topic/1022/what-are-you-doing-right-now/30339
-
@Dashrender said in What Are You Doing Right Now:
But I did infer (probably wrongly) that a single ARM would be single core, and now you're talking about 16x the overhead for 16 versions of XS running to support all of those processors.
Even the ARMs in cell phones are typically quad core if not octacore. My NVidia Shield TV unit is octacore, for example. Twice the cores of my Intel i7 laptop.
-
I think that efforts like what HPE and Dell have been doing to get very high density, very low cost ARM "blades" into the datacenter is where it is going to be. And that there will either be loads of hypervisor management used to push out workloads or something similar that does this for bare metal much like Ubuntu's MaaS (Metal as a Service) to get the ARM blades to act as much like VM hosts as possible.
