By: Brett (ggtgp.delete@this.yahoo.com), July 18, 2022 12:38 pm
Room: Moderated Discussions
Ben T (ben.delete@this.noemail.com) on July 16, 2022 4:50 pm wrote:
> >> Would you say even the Last Level Cache or System Level
> Cache should not be shared with an untrusted user?
> >
> > Yes, certainly. And not only that. Even memory and disks
> > shouldn't be shared. IMHO, even SAN shouldn't be shared.
>
> Do people here think the ideal public cloud infrastructure would be a cluster in a box rather than a single
> symmetric multiprocessor running virtual machines from multiple users? For example, imagine a 1U chassis containing
> 12 processor chips and 12 hot-swap 2.5” U.2 NVMe drives. Each processor chip could include:
>
> 8 performance CPU cores plus caches
> interfaces to two 128-bit LPDDR DRAMs
> 4 PCIe lanes to connect to one NVMe drive
> a 10G ethernet connection
>
> If the minimum unit a cloud user could get is 8 cores, they would not be sharing this chip
> with anyone else so the risk of data leakage between users would be greatly reduced. If
> it is possible to combine these processor units for bigger single tasks but still keep
> them partitionable into independent units for small tasks, that would be even better.
A 12 socket cluster in a box is far better than doing a 12 socket unified memory system.
Linux would be allocating memory two sockets away on the mesh and performance would go strait to hell. You would need a ridiculous number of PCIe channels in a mesh to try and overcome the bottlenecks by brute force, and pay a needless heat cost for that connect.
With cluster in a box you can use desktop/laptop chips and at most add a network switch to both share the networking and communicate between the sockets without hitting the Ethernet.
This is the future for 1U server boxes. But Intel and AMD have financial and competitive reasons to keep severs using DIMM’s and unified memory, moat maintenance reasons.
I expect the new server entrants to copy Apple with RAM on socket over the next 2-5 years, and only after loosing market share for Intel and AMD to release such sockets.
> >> Would you say even the Last Level Cache or System Level
> Cache should not be shared with an untrusted user?
> >
> > Yes, certainly. And not only that. Even memory and disks
> > shouldn't be shared. IMHO, even SAN shouldn't be shared.
>
> Do people here think the ideal public cloud infrastructure would be a cluster in a box rather than a single
> symmetric multiprocessor running virtual machines from multiple users? For example, imagine a 1U chassis containing
> 12 processor chips and 12 hot-swap 2.5” U.2 NVMe drives. Each processor chip could include:
>
> 8 performance CPU cores plus caches
> interfaces to two 128-bit LPDDR DRAMs
> 4 PCIe lanes to connect to one NVMe drive
> a 10G ethernet connection
>
> If the minimum unit a cloud user could get is 8 cores, they would not be sharing this chip
> with anyone else so the risk of data leakage between users would be greatly reduced. If
> it is possible to combine these processor units for bigger single tasks but still keep
> them partitionable into independent units for small tasks, that would be even better.
A 12 socket cluster in a box is far better than doing a 12 socket unified memory system.
Linux would be allocating memory two sockets away on the mesh and performance would go strait to hell. You would need a ridiculous number of PCIe channels in a mesh to try and overcome the bottlenecks by brute force, and pay a needless heat cost for that connect.
With cluster in a box you can use desktop/laptop chips and at most add a network switch to both share the networking and communicate between the sockets without hitting the Ethernet.
This is the future for 1U server boxes. But Intel and AMD have financial and competitive reasons to keep severs using DIMM’s and unified memory, moat maintenance reasons.
I expect the new server entrants to copy Apple with RAM on socket over the next 2-5 years, and only after loosing market share for Intel and AMD to release such sockets.
Topic | Posted By | Date |
---|---|---|
Retbleed | anonymous2 | 2022/07/13 03:14 PM |
Retbleed | anon2 | 2022/07/13 10:03 PM |
Retbleed | Adrian | 2022/07/14 12:05 AM |
Retbleed | Anon4 | 2022/07/14 02:17 PM |
Retbleed | anon2 | 2022/07/14 04:29 PM |
Retbleed | Anon4 | 2022/07/14 05:05 PM |
Retbleed | anon2 | 2022/07/14 05:37 PM |
Retbleed | anon2 | 2022/07/14 06:40 PM |
Retbleed | dmcq | 2022/07/15 04:54 AM |
Retbleed | anon2 | 2022/07/17 07:17 AM |
Retbleed | Michael S | 2022/07/15 07:08 AM |
Retbleed | Ben T | 2022/07/16 05:06 AM |
Retbleed | Michael S | 2022/07/16 11:41 AM |
Public cloud infrastructure | Ben T | 2022/07/16 04:50 PM |
Public cloud infrastructure | Rayla | 2022/07/16 09:15 PM |
Public cloud infrastructure | me | 2022/07/17 09:19 AM |
Public cloud infrastructure | Brett | 2022/07/18 12:38 PM |
Public cloud infrastructure | Adrian | 2022/07/18 01:19 PM |
Public cloud infrastructure | me | 2022/07/18 03:54 PM |
Public cloud infrastructure | Brett | 2022/07/20 03:35 PM |
Public cloud infrastructure | Brett | 2022/07/21 01:18 PM |
Public cloud infrastructure | inthestratosphere | 2022/07/21 02:46 PM |
Public cloud infrastructure | Brett | 2022/07/21 10:38 PM |
What’s needed for a viable Apple server? | Ben T | 2022/07/22 05:31 AM |
What’s needed for a viable Apple server? | Michael S | 2022/07/22 09:09 AM |
More DRAM capacity? | Mark Roulo | 2022/07/22 09:48 AM |
More DRAM capacity? | Doug S | 2022/07/22 11:05 AM |
More DRAM capacity? | Mark Roulo | 2022/07/22 11:20 AM |
More DRAM capacity? | Doug S | 2022/07/22 01:48 PM |
More DRAM capacity? | Wes Felter | 2022/07/22 04:49 PM |
Public cloud infrastructure | anon2 | 2022/07/18 04:25 PM |
Putting 12 processor packages in a 1U server | Ben T | 2022/07/22 10:02 PM |
Putting 12 processor packages in a 1U server | rwessel | 2022/07/23 07:15 AM |
Putting 12 processor packages in a 1U server | Daniel B | 2022/07/23 04:15 PM |
Putting 12 processor packages in a 1U server | Ben T | 2022/07/24 05:29 AM |
Multi-system cluster design space | Paul A. Clayton | 2022/07/24 08:49 AM |
Retbleed | Anon4 | 2022/07/15 03:00 AM |
Retbleed | Michael S | 2022/07/15 06:59 AM |
Retbleed | --- | 2022/07/15 11:14 AM |