Mike R.
asked on
Any real value to high IOPS disks (SDD vs. spinning) when serving network file systems (NFS/CIFS) across 10Gb
We're looking for a new server solution for our NFS/CIFS (only) file server(s). The network infrastructure will be 10Gb. Less than 100 users are connecting to the NFS/CIFS at any given moment. The files are a mixture of large and small (I.e. home directories, as well as large scientific data files in excess of 100GB.) The read/write ratio is probably around 70/30.
We are examining solutions like HP's 3PAR system, and Supermicro's 6048R-E1CR36L. Both are SANs with 12Gb backbones.
We are running ONLY file services on these servers (no virtual machine root virtual disks, no databases, no SaaS, etc.)
The only unanswered question is; what is the legitimate value of high disk IOPS in this environment and use case. SSD disks purport up to 75000 IOPS (as an array) , while the spinning disks up to 6000 (as an array).
My initial instinct is that SSD data drives have little to no value. But this is not supported by any facts.
Can anyone offer guidance?
We are examining solutions like HP's 3PAR system, and Supermicro's 6048R-E1CR36L. Both are SANs with 12Gb backbones.
We are running ONLY file services on these servers (no virtual machine root virtual disks, no databases, no SaaS, etc.)
The only unanswered question is; what is the legitimate value of high disk IOPS in this environment and use case. SSD disks purport up to 75000 IOPS (as an array) , while the spinning disks up to 6000 (as an array).
My initial instinct is that SSD data drives have little to no value. But this is not supported by any facts.
Can anyone offer guidance?
ASKER CERTIFIED SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
SOLUTION
membership
This solution is only available to members.
To access this solution, you must be a member of Experts Exchange.
Given the fact that there are multiple bits stacked on the platters it's not that hard and we're seeing consistent performance across 7200 RPM drives due to their ultra-high densities.
It's in the ballpark for max sustained transfer rate, Seagate claim 215MBps for the Exos 7E8 7.2K ones and Tosh claim 248 MiB/s for the MG07ACA. Wouldn't expect to see that in the real world though as there's always some seeks even if it's mainly sequential. There again internal flash and controller cache can boost that sometimes. One day they may even be faster than tape.
ASKER
@All
@Philip Elder
Thanks for the input. @Philip suggested running baselines on the existing server. I have two questions regarding that.
1. Can you suggest tools for this (to determine the actual IOPS versus read/write bandwidth). I have looked into IOMETER (running dynamo on the server) however this appears to test the drives directly using read/write algorithms to the drive hardware directly.
Since this system is running ZFS datasets, which the IOMETER/DYNAMO pair does not seem to recognize, I'm concerned it will screw up the dataset metadata and (basically) completely down the ZFS :D
2. How significant is the current IOPS to the needed IOPS?
Again in this scenario (which I did not fully describe - apologies) ... the current server is 5+ years old, running on a 1Gb (One gigabit) network. It is heavily overloaded, and causing data dropouts to users on the network (part of the reason for an upgrade).
The new server will be running on a 10Gb (ten gigabit) network.
Isn't measuring the current system somewhat like measuring a Ford Pinto in downtown traffic, to determine what you need for your undetermined new car that will drive on the autobahn? Shouldn't we choose the IOPS of the new server based on the minimum and maximum IOPS for its desired task?
I.e.
For file services: IOPS of 3000-10000* (HDD only)
For VMs: IOPS of 5000-25000* (SDD/HDD hybrid)
For video streaming, real-time data: IOPS of 10000-75000* (SDD only)
....that sort of thing?
* These numbers pulled out of the air, since this is the information I don't know :D
@Philip Elder
Thanks for the input. @Philip suggested running baselines on the existing server. I have two questions regarding that.
1. Can you suggest tools for this (to determine the actual IOPS versus read/write bandwidth). I have looked into IOMETER (running dynamo on the server) however this appears to test the drives directly using read/write algorithms to the drive hardware directly.
Since this system is running ZFS datasets, which the IOMETER/DYNAMO pair does not seem to recognize, I'm concerned it will screw up the dataset metadata and (basically) completely down the ZFS :D
2. How significant is the current IOPS to the needed IOPS?
Again in this scenario (which I did not fully describe - apologies) ... the current server is 5+ years old, running on a 1Gb (One gigabit) network. It is heavily overloaded, and causing data dropouts to users on the network (part of the reason for an upgrade).
The new server will be running on a 10Gb (ten gigabit) network.
Isn't measuring the current system somewhat like measuring a Ford Pinto in downtown traffic, to determine what you need for your undetermined new car that will drive on the autobahn? Shouldn't we choose the IOPS of the new server based on the minimum and maximum IOPS for its desired task?
I.e.
For file services: IOPS of 3000-10000* (HDD only)
For VMs: IOPS of 5000-25000* (SDD/HDD hybrid)
For video streaming, real-time data: IOPS of 10000-75000* (SDD only)
....that sort of thing?
* These numbers pulled out of the air, since this is the information I don't know :D
File services would be roughly 50% sequential / 50% random depending on file size.
VMs (unless only one VM running as a video server) are totally random.
Video is much more sequential, they even make disks specifically tailored to it. Compare WD Red Pro with WD Purple for example*.
*Actually they are tuned for multi-stream sequential writes rather than reads. Manufacturer doesn't even mention IOPS because that's not what they are for.
VMs (unless only one VM running as a video server) are totally random.
Video is much more sequential, they even make disks specifically tailored to it. Compare WD Red Pro with WD Purple for example*.
*Actually they are tuned for multi-stream sequential writes rather than reads. Manufacturer doesn't even mention IOPS because that's not what they are for.
I would like to see the test conditions for that!