Flashpool Configuration | NetApp | Page 1

torpid grove Jul 10, 2024, 4:58 PM

#

If I have a DS-224 Shelf of SSDs for setting up a Flashpool aggregate, how many disks should be used for setting up one hybrid aggregate per node?

mortal lichen Jul 10, 2024, 5:04 PM

#

The short answer is: it depends. Do you use sas or nl-sas for data? How many, how large? What is your active data footprint? What kind of io: random or sequential?

steep galleon Jul 10, 2024, 5:27 PM

#

Also: If you know which workloads you need to prioritize - so which actually need the performance and which not, most often it makes more sense to create an all-SSD aggr and no flashpool aggr. There are many situations where FlashPool does not really help. It only accelerates random reads and only random overwrites. I would check how high is your cacheable reads/writes number and then decide.

#

To determine the FlashPool cache size: https://docs.netapp.com/us-en/ontap/disks-aggregates/determine-flash-pool-candidacy-cache-size-task.html

Determine Flash Pool candidacy and optimal cache size

Before converting an existing local tier (aggregate) to a Flash Pool local tier, you can determine whether the local tier is I/O bound and the best Flash Poo...

torpid grove Jul 10, 2024, 6:01 PM

#

FAS8700 HA Pair using all 10TB NL-SAS disks except for the one shelf of DS-224 1TB SSD's. The entire workload is NFS presented volumes that are assigned to VMWare Datastores. Each volume is 80TB with about 50% usage max. Seems like there is not enough storage to create an SSD tier with only 1 SSD shelf of 24 drives that were purchased as a planned flashpool.

#

Also running 8 volumes per node.

mortal lichen Jul 10, 2024, 6:33 PM

#

How many VMs per volume? To be honest: this sounds like latency waiting to happen.

torpid grove Jul 10, 2024, 6:42 PM

#

Latency is between 1ms and 10ms on most of the volumes. They are not overloaded with VM's. 10 - 20 VM's per volume

#

Using capacity disks was not my choice. I inherited it.😆

mortal lichen Jul 10, 2024, 6:57 PM

#

That is not too bad. With the size of the volumes I was afraid of 40-50 VMs per volume.

#

As stated above make the choice for a SSD aggregate for performance VMs or a bit more performance for everyone.

torpid grove Jul 10, 2024, 7:02 PM

#

Thanks for your advice...

orchid rock Jul 11, 2024, 11:40 AM

#

one thing everyone forgets about the hybrid is the availability of in-line efficiency operations (without setting the nodeshell option).

steep galleon Jul 11, 2024, 12:35 PM

#

please elaborate

orchid rock Jul 11, 2024, 12:49 PM

#

if you have spinning media-only, the only way to enable in-line is to go modify a node option "sis.idedup_allow_non_aff_hya" (on each node). This is NOT recommended as it can KILL performance.

However, if you create a hybrid-aggr, you do NOT need to set the option and you can in fact enable in-line efficiencies
(see: https://docs.netapp.com/us-en/ontap/volumes/enable-inline-data-compaction-fas-systems-task.html)

The reasoning is that inline requires a LOT of meta-data operations...exactly what the SSD cache is used for. All the inline efficincies will happen on the cache/SSD

#

I have done this a lot with FAS/Hybrid/VMware-NFS

mortal lichen Jul 11, 2024, 2:48 PM

#

Thanks @orchid rock Great addition!

steep galleon Jul 11, 2024, 4:10 PM

#

But according to the write-cache policy only random-overwrites (and only if they are blocksize 16KB or smaller) will be sent directly to the Flash Pool SSD cache. So only some of the writes would get the inline-dedup benefit.
Or does enabling this option changes that so that all writes go the SSD cache?

orchid rock Jul 11, 2024, 4:34 PM

#

Not entirely sure, I just know it works. Have not “dug deep” into it

dapper kelp Jul 11, 2024, 4:44 PM

#

FlashPool is more for read caching of random reads. HDDs are better at sequential performance than random.

#

You need to know if your workload is random or sequential first. I think there is an analyzer (AWA?) the account team can work on to see what the workload is like.

#

Honestly if you can swing it, C series is much better.

#

What I see is a lot of customers buy FlashPool thinking the cache will take the hit in load off of disks, but either it's a sequential workload so it doesn't get cached, or the CPU overhead of FlashPool overwhelms CPU. FlashPool can be CPU intensive compared to non-FlashPool.

dapper kelp Jul 11, 2024, 4:47 PM

#

steep galleon To determine the FlashPool cache size: https://docs.netapp.com/us-en/ontap/disks...

Yeah it's AWA. OG1 linked the documentation here.

calm crag Jul 11, 2024, 5:16 PM

#

Sadly, FabricPool is not possible on FlashPool Aggregates (and there's no way to remove the FlashPool part after it has been added, except deleting the aggregate).
Also, at least in the past we got bitten by lots of nasty bugs and issues with FlashPool, that is why we haven't installed any FlashPool in over 4 years or so now.
A small C-Series with a FAS for FabricPool is often the neater solution

#Flashpool Configuration