Nimble H/W controller specification

[Beegad]

Kindly i would like to ask, what is the specs of nimble controller? internal Processor and internal cache?

[Nick_Dyer · Accepted solution]

Hello,

We do not publically post the specifications of the Nimble models, as we do not associate a certain value or technical benefit to the amount of CPU or RAM cache directly available inside the controllers. Historically, we've seen customers attempt to compare controller specs in order to obtain value for money - which isn't how enterprise storage works.

For example:

Nimble might be 1x Intel Gold CPUs, 64GB RAM

Competitor A might be 2x Platinum CPUs, 512GB RAM

You would think based on the above specs, the competitor might perform better as you're getting more kit for your money...

...however then you realise, after diving deeper, the performance would be the same or even better on Nimble vs the competitor - it makes the analysis a red herring.

This is because the competitor would be less efficent, and have higher overheads of it's system processes - meaning it NEEDS that amount of resource in order just to keep 'the lights on' as it were.

Hope this helps!

[Beegad]

Many thanks Nick for your quick reply, i got the idea

[NomadeWolf]

"Nimble utilises caching technology to write hot data straight to the SSD layer, before stripping the data to lower-cost, large-format nearline disks. Unlike traditional arrays, the spinning disk and SSD components do not directly add performance to the array, meaning Nimble arrays drive their performance directly from the CPU and RAM capabilities of the controllers."

Taken from https://www.checkmark-it.com/news/hpe-nimble-vs-hpe-3par-the-key-differences 

[plasticwax]

It would help us understand why a 5 year old 3par is out performing your HF40C when all we are doing is copying files to it as pert of our migration, but continue to see latency and poor performance.

As some one certified in EMC and NetApp storage, I have heard some amazing stories about Nimble... but from my opinion its not even on the same planet as these two.

1. CASL seems to dictate that everything is done in ram? specifically... compression... yet... it appears as we copy data over that is not compressed the memory does not seem to be doing anything? while we get alerts about CPU performance affecting write latency? the array is barely 5% used... som specs around the CPU would help us understand if you put a crappy celeron in it or not.

Stats from 2020-06-25,12:38:58 to 2038-01-19,14:14:07
vol reads rdKiB writes wrKiB seqRead% nsMemHit% nsSsdHit% dataKiB dataSnapKiB
volname 74 296 5 992 0 0 100 0 0
volname 67 267 11 1376 0 1 98 0 0
volname 50 224 19 1020 0 0 100 0 0
volname 49 196 11 560 0 0 100 0 0
volname 72 288 101 2304 0 0 100 0 0
volname 62 248 0 0 0 0 100 0 0
volname 68 272 7 1240 0 0 100 0 0
volname 80 320 5 1572 0 0 100 0 0
volname 75 300 11 1630 0 0 100 0 0
volname 71 308 5 980 0 0 100 0 0
volname 68 272 6 1332 0 0 100 0 0

Please advise why the ssdhit rate is sitting at 100%, while the memory is literally doing nothing and we are constantly writing which means the compression should be constantly going... is compression really done in DRAM or CPU also?

thanks

[Nick_Dyer]

Hello and good day,

yes all IO is compressed in DRAM before being sequentialised to disks. Unless you turn compression off on a volume or it's performance policy, that is.

The STATS command you're running in the array CLI is reporting the cache hit for reads from either SSD or memory (CASL services random read IO from NVDIMM, DRAM and SSD). Writes never get serviced through SSD. The command you're running doesn't show the utilisation of components of the array for interactions such as compression.

If you haven't yet, please open a support case with Nimble Support so we can investigate further. One thing I highly recommend doing is a 'plumbing' test with VDBench to test the maximums of the array, as well as to ensure that all configuration is correct through your stack. Copying files is not an adequate performance test for real-world workloads, as it is a single-threaded workload.

The HF40x has 16 cores of Intel CPU per controller, it's certainly NOT a Celeron.

[JesusHotburger]

HPE may not list their CPUs but don't be fooled by the downplaying of CPU.  I just had to upgrade my controllers because the CPU was totally overworked and could not keep up with the workload.  I agree with the analogy they don't have to be Platinum processors but in the Nimble world properly sizing the CPU i.e. model of the controller is extremely important.