L3 cache importance

[Robstarusa]

Hello,

I am going to be running a 4 server cluster with shared NFS filesystem (possibly on one node, possibly on a 5th node...).

I have some quite old cpu's, except for one I traded for:

Currently only "new" cpu: athlon ii x4 620
Old: be-2350, amd 4850e, amd 5050e.

Proposed new processors:
Either amd 910e (~ $190 USD) or 605e ( ~ $155)

My other thought was to get something like 2 405e's and one 965 (and keep it underclocked/undervolted).

I cannot find linux/kvm benchmarks comparing these two. Is it worth it to pay a premium for phenom ii or should I stick with athlon ii? I prefer the low power usage if possible.

 

[udo]

Hi,
what kind of benchmark do you need/want?
I have the 965, 620 and 4850e in easy access.

Udo

 

[Robstarusa]

I'd like to see how say, the 965 reacts vs the 620 when trying to run, say 16-32 vms' on 8G ram. Is the 965 able to do it at a decent speed where the 620 is not?

How is the performance of the vm's under some load?

 

[udo]

I'd like to see how say, the 965 reacts vs the 620 when trying to run, say 16-32 vms' on 8G ram. Is the 965 able to do it at a decent speed where the 620 is not?

How is the performance of the vm's under some load?

Hi,
i think such an test is impossible. It's depends highly of the vm (openvz/kvm and what usage).
Imho are 8G ram for 16-32 vm's pretty insufficient. On our two produktionservers are app. nine vms on each node (kvm) and we have 16GB. But i have reserves (to get all vms on one node, if i had trouble with one). Some more vms are ok but 16 not.
Most important than cpu-power is the IO - good raid with fast sas-disk, good NIC, perhaps FC-San makes more differents as the switch from 620 to 965 (imho).
The motherboards has ddr3-ecc-ram - this gives also an performance benefit (ddr3 not ecc - but i don't know exacty how much).

Udo

 

[Robstarusa]

I'm using only KVM (as I use more than just linux) with the new kernel with KSM.

My proxmox cluster will be made up of multiple SSD's (Intel/OCZ) over nfs with bonded gigabit ethernet.
Backups will happen nightly to a repo of

My vm's (to start with) will consist of

3 internal nameservers (probably 256 ram ea.)
1 openbsd firewall (256 ram is plenty)
1 local ubuntu repository (512 ram)
2 public authoritative dns servers (256 ram ea) - non recursive
1 public non-auth recursive nameserver (256)
1-4 desktop machines with different distros to play with (perhaps 512M-1G ram ea)
1 asterisk server (256 ram -- no transcoding)
1 vpn server (256 ram)
1 very light web/db server (512 ram)

That is 15 boxes with under 8G ram before considering KSM memory reduction

About 6 of those currently exist & I see 50-75% cpu usage on the host (of 1 core) all the time...I'd like to reduce this significantly.

 

[udo]

I'm using only KVM (as I use more than just linux) with the new kernel with KSM.

My proxmox cluster will be made up of multiple SSD's (Intel/OCZ) over nfs with bonded gigabit ethernet.
Backups will happen nightly to a repo of

My vm's (to start with) will consist of

3 internal nameservers (probably 256 ram ea.)
1 openbsd firewall (256 ram is plenty)
1 local ubuntu repository (512 ram)
2 public authoritative dns servers (256 ram ea) - non recursive
1 public non-auth recursive nameserver (256)
1-4 desktop machines with different distros to play with (perhaps 512M-1G ram ea)
1 asterisk server (256 ram -- no transcoding)
1 vpn server (256 ram)
1 very light web/db server (512 ram)

That is 15 boxes with under 8G ram before considering KSM memory reduction

About 6 of those currently exist & I see 50-75% cpu usage on the host (of 1 core) all the time...I'd like to reduce this significantly.

Hi,
your vms have a small footprint - so perhaps you are lucky with "only" 8GB Ram.
Do you think it's the right way to use the fast ssd via nfs? nfs is not realy known for top-speed, but i'm using nfs only for sharing home-dirs. I think iscsi was a better solution but my iscsi-tests are also not the best (i stay on FC-SAN - perhaps it's change with 10G-ethernet).

I quite agree with you, that 50-75% constant cpu-usage is too much. But the different in price between a 620 and 965 are not to much - so i will take the 965. I advise a quad (or six) core, because of the better virtualisation support - it's possible to use hardware-virtualisation in a kvm-guest.

Udo

 

[Robstarusa]

the 620/965 dont' have much price difference, but I'm really interested in the 45W quad cores. I have 4 boxes, and the "server room" is in a basement with the furnace that heats the house. With 3 45W cpu's right now it is around 70-80F (21-28C I think).

I'd really like to keep the power usage low(er) if possible.

 

[udo]

the 620/965 dont' have much price difference, but I'm really interested in the 45W quad cores. I have 4 boxes, and the "server room" is in a basement with the furnace that heats the house. With 3 45W cpu's right now it is around 70-80F (21-28C I think).

I'd really like to keep the power usage low(er) if possible.

Hi,
you are right to save energy. But if you get a 965, you have also the cpu-power if you need them.
I measure the overall power-consumption of one of the proxmox-server:
round 258 Watt without significant load up to 313 Watt at full load (4 kvms are gzip /dev/null and a pveperf to make driveusage).
I think it's not too bad, for a system with 16GB Ram, Raidcontroller, 2 NICs, FC-Controller, 2 Sata-Disks and four SAS-Disk and the Dual-Powersupply.

To compare this with another CPU isn't easy, i have the same motherbord, but not the same disks to test.
Perhaps you get the same powerconsumption as the old boxes with a bigger cpu due to better (in meaning of power saving) disks,motherbord and powersupply?

Udo

 

[Robstarusa]

Very interesting. I am pretty sure my 5050e/4850e/2350 are all around 100W under load ea.

Also: What coolNQuiet steppings do the 965/955 support?