Effects of pairing 7900X with 192GiB DDR5 ECC memory

Hardware Hub Build a PC
, , ,
1

I have a TrueNAS Scale based NAS which has a 7900X on a X870 ProArt motherboard with two sticks of 48GiB DDR5 ECC UDIMM. For storage I have a Broadcom 9305 - 16i with Adaptec 85255
For context, the NAS is serving my 4 Proxmox nodes over iSCSI and/or NFS shared storage (depoending on latency sensitivity of the application the VM is serving) from twelve 4 TB SATA SSDs in a 4-wide stripe of RAID-Z1 and files over CIFS from four 4TB 7200 rpm SATA HDDs in RAID-Z2. The connection to the Proxmox nodes is over a bonded triplet of 10GbE links
It’s not being used to serve VMs from the machine itself.

I’m thinking of upgrading to four sticks of 48GiB by doubling the number of sticks with another pair of the same kit.

I am slightly worried that this might affect performance in peak usage scenarios. Assuming it will run, I am still unsure about:

  1. How much will the 3600MT/s transfer rate, combined with DDR5’s higher latencies, affect IPC on the 7900X? Will the reduction in performance be noticeable for the transfer rate for the iSCSI transfers?
  2. If I enable Eco mode on the CPU, will this further affect IPC noticeably?

Are there any other things I need to look out for or think about before executing this upgrade?

1 Like
2

What does that mean, 2x3 ports?
Is this a single user NAS or multiple clients at the same time?

3

Is LACP bond over 3 ports a better description?

4

What if you set the memory speed on your current setup to 3600MT/s and benchmark it before/after?

5

If this is a storage box, your network will be the bottleneck, then the disks, then the LSI card, then the RAM, then the CPU. Even with bonded 30gbit, ram quantity will matter more than the speed. (And you can run it above 3600 if you really want to most likely, but as said, there is not much of a point to it if this is a storage appliance.)

1 Like
6

I just want to be sure that we talk about LACP, because I was told in my last Cisco training, that the hashing algorithm of LACP works best with 2, 4, or 8 links.
The modulo operation distributes traffic evenly since these numbers align well with binary hashing mechanisms (powers of 2).

Beyond that, I think like @Strawberry, that with a little patience and dedication to the topic, you’ll get somewhere between 4800 and 5200 MT/s, but in any case, you won’t see any meaningful difference.

1 Like
7

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.