Gen6 A200/2000 node: what can you expect ?

Hello everybody,

[I've also posted this message on the Isilon Technical User Google Groups]

Gen6 nodes are new and I understand that it is gonna be difficult to have feedback from the field.

But I'll try. Our technical contact told us there is already some A2000 cluster in the U.S, so maybe

we could have something interesting here ! Of course our different Isilon contacts are involved, but

we are interested in all the feedback we can get.

Let's think about an Isilon cluster with 30% of H500 nodes and 70% of A200/2000 nodes. I've got 70%

of cold data that I can identify and move to the A2xxx nodes using a Smartpool policy. Everything

looked fine until I saw some of the hardware for the A nodes: a 129$ D1508 2 cores (4 HT) Pentium

Processor and 16 GB of ram.

What do you think of this in a scenario where only H nodes will handle external cifs/nfs clients

(A2xxx nodes will only be using the internal - new - ethernet backend), 99% of the activity will

happen on the H500 nodes with 1% (or less) of traffic to the A2xxx nodes ? Even with this quantity

of RAM and this CPU, I guess it should work quite well.

My concerns are more about the different jobs running on the cluster, like fsnalyze, multiscan,

snapshot deletion, synciq... I don't have enough knowledge how the job engine is working, but I

understand that most of the job will involve all the nodes or most of them.

What will be the impact of such a difference of computing power between H500 and A2xxx nodes in the

jobs runtime ? What about the clients side of thing ? Will everything be throttled down ? Are we

talking order of magnitude here or something more like 10/20% slower ?

We had a lot a problems (fsanalyze taking weeks, multiscan always failing, snapshot deletion taking

a lot of cpu for long period of time...) with the job engine (S200/X400/NL400). We never really had

definitive answers about that besides upgrading OneFS. There has been a lot of improvement in OneFS

8 (fsanalyze using snapshot differences is one them and seems to work quite well after the initial

scan) but I can't stop myself that all of that could have been caused by hardware difference between

nodes. Maybe or maybe not.

Our NL400 have 48 GB of memory, two cpu with 4 cores (no HT) each. That's a lot different from 2

cores and 16 GB of ram and I'm really wondering what could be the side effects in the ideal scenario

I've mentionned).

While I'm here, what will you choose between L3 cache and GNA (if we have enough SSD space to

activate it of course) in a mixed H500/A2xxx cluster ?

Jean-Baptiste