Epyc to replace E5-2630

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
J

Jamy

Member
Mar 29, 2017
44
6
8
48
Hi all,

I currently have a XCP-NG machine with dual E5-2630 v3 processors at 2.4 ghz, with 96GB of RAM, and dual 10g NIC. My VM's are frequently not very responsive. I'm looking at possible upgrades. Would an Epyc 7542 make a difference? I haven't used Epyc, however there seem to be decent deals on them. Any advice would be appreciated.

Thanks
 
T_Minus

T_Minus

Build. Break. Fix. Repeat
Feb 15, 2015
7,762
2,141
113
This won't be a big jump... but what about going down to a single V4 like a 1650 V4? or You could also go to Dual 2667 V4.
Those are very cheap CPUs right now, and won't require a rebuild.

I wouldn't go to that epyc for better snappyness single core performance, I'd be looking at the latest generation or even 5950x if you want to get really snappy.
 
J

Jamy

Member
Mar 29, 2017
44
6
8
48
Last edited:
J

Jamy

Member
Mar 29, 2017
44
6
8
48
T_Minus said:
This won't be a big jump... but what about going down to a single V4 like a 1650 V4? or You could also go to Dual 2667 V4.
Those are very cheap CPUs right now, and won't require a rebuild.

I wouldn't go to that epyc for better snappyness single core performance, I'd be looking at the latest generation or even 5950x if you want to get really snappy.
Click to expand...
I was hoping to be able at least re-use my 96GB of DDR4 ECC.
 
P

Pete.S.

Member
Feb 6, 2019
56
24
8
96GB is a bit of an odd configuration for that E5 v3/v4 CPUs. For maximum performance you should run a balanced memory config on the 4 memory channels and ideally have max 2 DIMMs per channel. That means 4 or 8 DIMMs per CPU.

In practical terms using 16GB or 32GB DIMMs with 2 CPUs you end up with:
  • 128GB (2x4x16GB)
  • 256GB (2x8x16GB or 2x4x32GB)
  • 512GB (2x8x32GB).
But RAM bandwidth is not usually the big bottleneck. My guess is that you can got a lot more performance out of what you have with small changes. Both on the hardware side and on the host / guest configs.

What VMs are you running? What disks do you have? Are you sure you're constrained by CPU and not something else?
 
  • Like
Reactions: ano
P

Pete.S.

Member
Feb 6, 2019
56
24
8
If I wanted to upgrade the CPUs I'd use swap in two E5-2680 V4.

It's better for virtualization because it has more cache, more cores and still better single core performance compared to e5-2630 v3.
And it's ridicously cheap - $30 or so per CPU.
 
V

vvkvvk

Member
Feb 1, 2023
90
36
18
Just throwing in nice spec v4s and balance out the RAM would be my choice.

Epyc (Rome) is just not that impressive compared to the best of what Broadwell-E has to offer and you can run any garbo tier slow RAM and LR DIMMs on it. I went from twin 99v4s to cheap Rome and it is sort of stalemate when it comes to perf, however, the upgrade path is vastly superior with Milan and Milan-X.
 
I

i386

Well-Known Member
Mar 18, 2016
4,411
1,638
113
35
Germany
Jamy said:
My VM's are frequently not very responsive.
Click to expand...
I have some single socket e5 v3 and v4 and they are still pretty powerfull for many applications. If I were you I would investigate what makes the vms "not very responsive" and depending on the results there could be many other options (higher clocks v3/4 or more cores, faster and more pcie lanes etc.)
 
  • Like
Reactions: T_Minus
J

Jamy

Member
Mar 29, 2017
44
6
8
48
Pete.S. said:
96GB is a bit of an odd configuration for that E5 v3/v4 CPUs. For maximum performance you should run a balanced memory config on the 4 memory channels and ideally have max 2 DIMMs per channel. That means 4 or 8 DIMMs per CPU.

In practical terms using 16GB or 32GB DIMMs with 2 CPUs you end up with:
  • 128GB (2x4x16GB)
  • 256GB (2x8x16GB or 2x4x32GB)
  • 512GB (2x8x32GB).
But RAM bandwidth is not usually the big bottleneck. My guess is that you can got a lot more performance out of what you have with small changes. Both on the hardware side and on the host / guest configs.

What VMs are you running? What disks do you have? Are you sure you're constrained by CPU and not something else?
Click to expand...
My currently running VM's are:

Security Onion (Linux) configured with 12gb RAM.
Wazuh (Linux) configured with 8gb RAM
Windows 10 with 8gb RAM
Xen Orchestrator - 3gb RAM
XCP-NG control domain uses 5.8gb of RAM

For Disks, I have 4.2 TB of spinning disks in a ZFS stripe array. Only 28% in use.

For Networking, I have two 10g Chelsio cards in a bonded pair.

The MB is a SuperMicro X10DAL-i

I've attached a performance graph. This is what it looks like pretty much all the time. I have never seen the CPU's spike to 100%
 

Attachments

P

Pete.S.

Member
Feb 6, 2019
56
24
8
vvkvvk said:
You just limit your VMs to use CPUs, memory and caches on certain node. I don't have faintest idea how to do it on xcp-ng / Xen as I haven't used it all, this is what Google got me and it looks fairly simple Compute and GPU | XCP-ng documentation
Click to expand...
I've ran a lot of tests and xcp-ng can handle numa architecture and put the right cores on the right CPU pretty good. So pinning cores has limited use. The only time I would consider it is if you dedicate cores at all times to a VM for latency reasons and accept overall performance loss. Or if you want maximum I/O performance on one VM so you pin it to the CPU that has the PCIe bus connection to the I/O device.

A bigger and perhaps more common problem is when VMs have too many vCPUs assigned to it. Then the host can't schedule the VM efficiently.
 
P

Pete.S.

Member
Feb 6, 2019
56
24
8
Jamy said:
My currently running VM's are:

Security Onion (Linux) configured with 12gb RAM.
Wazuh (Linux) configured with 8gb RAM
Windows 10 with 8gb RAM
Xen Orchestrator - 3gb RAM
XCP-NG control domain uses 5.8gb of RAM

For Disks, I have 4.2 TB of spinning disks in a ZFS stripe array. Only 28% in use.

For Networking, I have two 10g Chelsio cards in a bonded pair.

The MB is a SuperMicro X10DAL-i

I've attached a performance graph. This is what it looks like pretty much all the time. I have never seen the CPU's spike to 100%
Click to expand...
We run xcp-ng hosts at work and I don't see a problem with what you're running. Except ZFS and spinning disks. Those a red flags to me when it comes to the snappiness of the VMs.

You actually have a fair amount of CPU usuage for your few VMs. I'd take a look at individual VMs to see where the usage comes from.
 
  • Like
Reactions: T_Minus
A

ano

Well-Known Member
Nov 7, 2022
698
298
63
zfs + spinners are awesome, zfs is really good at this.

but. vms on spinners? very slow vs ssd
 
  • Like
Reactions: T_Minus
T_Minus

T_Minus

Build. Break. Fix. Repeat
Feb 15, 2015
7,762
2,141
113
Yesss, add NVME(s) for VM storage, and 2697 v3 or 2680 v4 for dirt cheap, and adjust RAM if needed.
 
D

drdepasquale

Active Member
Dec 1, 2022
103
37
28
Jamy said:
Hi all,

I currently have a XCP-NG machine with dual E5-2630 v3 processors at 2.4 ghz, with 96GB of RAM, and dual 10g NIC. My VM's are frequently not very responsive. I'm looking at possible upgrades. Would an Epyc 7542 make a difference? I haven't used Epyc, however there seem to be decent deals on them. Any advice would be appreciated.

Thanks
Click to expand...
Switching from Haswell Xeons to EPYC Rome is a major improvement in per core performance and energy efficiency. Combined with PCIe 4.0, faster 3200 MHz memory, a the 7nm process EPYC is well worth it.
 
You must log in or register to reply here.
Top Bottom