HPE Storage Users Group

Thanks Justin! That actually makes a huge difference in my calculus, if I can put a couple hundred GB of SLC for write caching plus a few TB of cMLC for warm data and take advantage of dedupe and then allow everything else to tier down to NL that would probably make a 7200 cost competitive with Dell. Now it'll just be down to timing, I can probably ignore SSD caching at implementation since high performance isn't a hard requirement in our case, just something we know we'll eventually need if we ever get into a real DR scenario.

Have they released any technical details as to how this will be implemented?

For instance, is it system wide, or on a per cpg basis?

The other oft overlooked benefit of dedupe is that it can seriously improve the performance of spinning disk by reducing seek time, and improving.the cache hit percentage.

Afidel,

Can you share some of the details of what dell is proposing? I would be very interested to see how they've gotten the price so low

I'm actually meeting with them next week, but the numbers I'm using are based on a friend in the industry that got one of their first all flash Compellent arrays for $3/GB usable with 5 years support (roughly what we paid for our initial buy of 3 tier 3par 7400, though with the recent capacity expansion we're well below that) that was their stock config with a few SLC drives to absorb writes and a large pool of eMLC for reads.

Afidel,

I will say in my experience, the cost of the 7200 is dramatically lower that the 7400.

When we bought our 7200's, we originally were looking at the 7400. Because of the way hp caps the per drive ltu's the 7200 was something like 75,000 dollars cheaper for the same spindle counts. Couple that with the fact that since 3.1.3 the 7200 supports up to 240 drives, you may be able to step it down and save some green.

Our two tier 7200s with 50 tb total usable, and 48 spindles on the fc tier cost us 2.40/gb, and almost all our capacity is on fc (we're expanding nl this year)

According to techopsguy the dedupe is per CPG, I wonder what that means for AO? Will it require an AO policy to contain only CPG's with it on or off?

As to 7400 vs 7200, we had to go 7400 as at the time 7200 was much too small and even today it is with traditional disks. We're at 65TB usable, much of it in FC and we grow at 50% per year compounded so a 7400-2/7200 won't be enough in 4 years using 900GB FC. That's why we recently upgraded to a 7400-4, the numbers pointed that we wouldn't be able to make it to the end of year 4 (let alone year 5) given where things were going and it was easier to expand now then try to do it with an almost full chassis. Since we only paid ~$1/GB with controllers, shelves, warranty, and install/data migration services for the expansion it was an easy decision to make. For our DR scenario a 7200 with SLC/cMLC/NL should be sufficient since we should be able to fit 300TB with a few dozen TB active into a such a config.

Indeed. You could put in 192 2tb drives, 24 cMLC, and 24 SLC

Though depending on your performance needs, you might be able to do just 48 cMLC and call it a day

Man, 30tb/year is a boat load of growth (for me). Our environment is more traditional. We only grow maybe 10 tb/year if that

The flash cache has limits, I do not know if HP disclosed them publicly yet, but it is not that large and it depends on the model array as to how large it is.

7450 only is purely a marketing decision, no one else has inline dedupe for spinning disk, so best to hit the all flash array startups that do this currently first.
3.1.3 MU1 Introduces cMLC support
3.1.4 Introduces thin dedupe and other stuff
SLC is a dead end eMLC & cMLC going forward, adaptive writes, sparing, zero detect, dedupe and other under the covers tech make SLC unnecessary and the speed difference is marginal in a system optimized for flash with the above features.
Dedupe is per CPG.
Dedupe is SSD only (for now).
Not relevant for AO as AO can use spinning disk (see above response).
If you're ingesting directly to SSD then what's the point in extending cache with SSD, your likely just slowing the write down. Speed to write to SSD disk tier or SSD cache will be very similar but won't need the additional cache flush.
From experience Dell can only actually undercut by not doing an apples for apples similar to EMC, the SLC/MLC tiering is a fix for their own architectural issues rather than a feature to benefit Customers.

Thanks for the info!

I'm curious as to why dedupe is SSD only. They say that the ASIC is doing the heavy lifting to make it happen. So wouldn't the storage medium be irrelevant at that point? Unless possibly they need to do a read from disk in order to make the dedupe work, in which case the spinning rust wouldn't be quick enough for dedupe to work well inline without a major performance hit.

Or maybe it's just a marketing thing to attack the other AFA's which have dedupe, and have been possibly costing HP sales in the AFA market, where in the spinning disk market nobody else has inline dedupe so not having it isn't costing HP any money.

The reason I keep asking about the dell specifics is that I cannot see how they would truly be so much less expensive apples to apples. Not long ago we collected proposals from HP and 2 other vendors selling the same hardware. Without disclosing too much. let's say HP came in at 200,000 (prices and names changed to protect the innocent ( and guilty ) ). Vendor A came in at 400,000 with an apples to apples quote. Vendor B came in with the same manufacturer as vendor A, but at 190,000. The difference was HP and Vendor A proposed 16 SSD's, 52 10K, and 24 7.2K (the performance characteristics were derived from an indepth survey of the environment, hence, the spindle count was important). Vendor B comes in with 12 SSD's, 16 10K, and like 52 7.2K. The usable capacity was there, but the only way they were able to undercut HP was to ignore the spec and propose an array with much lower performance characteristics

My guess is they're keeping the dedupe database on the admin section of the CPG to reduce the memory footprint. As an example of how memory intensive dedupe can be ZFS uses 300B per block for dedupe, for 16KB blocks (what 3Par uses) that works out to ~60GB of ram for 12TB of pre deduped blocks. You can see that supporting hundreds or thousands of TB of source data necessitates making a space/time tradeoff.

As to the Dell thing, it's not hard to come out below $8/GB of MLC flash, truly consumer grade drives are at $.80/GB and Samsung and Seagate enterprise class eMLC are at ~$1.10/GB at retail (Newegg pricing) so there's plenty of money to be made at $3/GB usable.