CPU overhead for multiple SMB clients

[DaShesano]

Hi all,

I recently set up a storage server with 8 4tb seagate exos drives and a samsung 883DCT as a read/write cache (with the 8 drives in parity) using tiered storage through storage spaces direct in Windows Server 2019. I was intending it to serve 8-12 gigabit clients concurrently, and was wondering how many CPU cores would be necessary in order to support all of the clients in a worst case scenario (50/50 reads/writes). I am currently using an EPYC 7281 and was hoping for some CPU resources left to run some VMs.

Thanks,

P.S 

On the topic of VMs, I was wondering what good GPUs you guys recommend, I am currently using Hyper-V and am considering either a tesla M4, or an M40 12gb as there's some compelling deals on ebay, I'm going to be doing some modeling work on the accelerated VMs.

[Windows7ge]

8 to 12 clients doing file transfers shouldn't cause very high CPU utilization. Your pool will become a bottleneck before the CPU. 4C/8T will probably suffice but if you have an EPYC 7281 that's definitely more than enough & then plenty for a series of VMs.

 

I can't comment on GPU's for pass-though. Not a field I know much about.

[Electronics Wizardy]

You won't need much at all. You will be disk and network limited well before the cpu limits you.

 

16 hours ago, DaShesano said:

I recently set up a storage server with 8 4tb seagate exos drives and a samsung 883DCT as a read/write cache

How is that cache going? Id use refs if you can as caching is much better in refs. Is it all mirrors?

16 hours ago, DaShesano said:

On the topic of VMs, I was wondering what good GPUs you guys recommend, I am currently using Hyper-V and am considering either a tesla M4, or an M40 12gb as there's some compelling deals on ebay, I'm going to be doing some modeling work on the accelerated VMs.

How much modeling horsepower do you need?

 

Are you running RDS? The software won't be cheap to use these vms, so for just a few vms, id just get anouther workstation and rdp into it.

 

 

[DaShesano]

1 hour ago, Electronics Wizardy said:

How is that cache going? Id use refs if you can as caching is much better in refs. Is it all mirrors?

I've just used the standard storage spaces setup, whatever file system that uses I have used. Its only set up in parity (so iirc 2 parity disks) across all the drives.

 

2 hours ago, Electronics Wizardy said:

How much modeling horsepower do you need?

 

Are you running RDS? The software won't be cheap to use these vms, so for just a few vms, id just get anouther workstation and rdp into it.

I've noticed that the performance is best when used with a 960, and after that I see no benefit. I was thinking that the use of the bigger tesla would let me run more simultaneous sessions. All of the software I'm using is through academic licenses or open source, including operating systems, so their cost is negligible, I was hoping to do this for a one box solution. 

[Electronics Wizardy]

3 minutes ago, DaShesano said:

I've just used the standard storage spaces setup, whatever file system that uses I have used. Its only set up in parity (so iirc 2 parity disks) across all the drives.

 

Use refs if you can, its much better with tiering in storage spaces.

 

Also parity has known performance problems with writes, id stay away.

4 minutes ago, DaShesano said:

I've noticed that the performance is best when used with a 960, and after that I see no benefit. I was thinking that the use of the bigger tesla would let me run more simultaneous sessions. All of the software I'm using is through academic licenses or open source, including operating systems, so their cost is negligible, I was hoping to do this for a one box solution. 

You got price estimates, a m4 is similar to a single 960, a m40 should be about double the speed as its simimlar to a 980ti. Look at m60's aswell.

 

Is this running through VDI in windows? Do you have VDI setup yet?

[DaShesano]

6 hours ago, Electronics Wizardy said:

Is this running through VDI in windows? Do you have VDI setup yet?

I was planning on running it through hyper-v and using either RemoteFX or DDA (I know that there's issues surrounding RemoteFX on server '19). So far I haven't set up any of the VMs yet, I was intending two to run Windows 10 and one possibly to run Ubuntu 18.04.

[Electronics Wizardy]

2 hours ago, DaShesano said:

I was planning on running it through hyper-v and using either RemoteFX or DDA (I know that there's issues surrounding RemoteFX on server '19). So far I haven't set up any of the VMs yet, I was intending two to run Windows 10 and one possibly to run Ubuntu 18.04.

remoteFX is removed from the gui in server 2019. Id test it on a workstation to see if it would work, cause then you could use a much cheaper consumer gpu.