An all-flash build of Hitachi's VSP array has claimed the second-highest SPC-1 benchmark, the highest for a traditional storage array. The SPC-1 benchmark tests the performance of storage arrays doing random read and write I/O operations typical of business-critical applications. An all-flash VSP, using Hitachi's Accelerated …
Almost as pointless as 100% cache hit benchmarks. Few people can afford all-flash arrays. They need to start benchmarking hybrid arrays with hybrid workloads.
I'd be interested to see IBM's response with the TMS arrays, but ultimately that will have little real-world value other than bragging rights and twisted false promises from salesmen.
IBM did test hybrid a few years ago, it's the only one I've seen to-date though.. the results weren't pretty.. just copy pasting from my blog post at the time
Configuration of the system included:
1 IBM DS8700
96 1TB SATA drives
16 146GB SSDs
Total ~100TB raw space
Performance of the system:
34.1 TB Usable space
Cost of the system:
$1.58 Million for the system
$47.92 per SPC-1 IOP
$46,545 per usable TB
Nice to see IBM was good enough to release them even if the results weren't pretty.
Yes, I was thinking the same. The TMS (IBM FlashSystem) array on that chart is ancient, a 620 model. The new 810s and 20s run at 500,000 IOPS random r/w sustained with sub-ms latency. Not DRAM performance, but it should easily be better than VSP and be top of the non-volatile arrays.
Most people only need to run their primary OLTP DBs in all Flash, which is probably a few TBs. Even if you are paying $15,000 per TB, it is not all that expensive.
The SPC-1 benchmark result you are referring to is at http://www.storageperformance.org/results/benchmark_results_spc1/#a00092
The purpose of that particular SPC-1 submission was to demonstrate the effectiveness of the DS8000 standard Easy Tier feature at reducing average I/O response time through intelligently and nondisruptively relocating sub-volume extents. Graphs in the detailed Full Disclosure report show this significant improvement occurred as expected. The benchmark was *not* designed to demonstrate maximum DS8700 performance. (By the way, the DS8700 is two generations behind the current DS8870 model which has a published SPC-1 result of about 451K IOPS for an all-HDD configuration.)
I suspect many people who observe SPC benchmark results do not realize that configurations vendors use for an SPC benchmark are not necessarily intended to demonstrate maximum performance a system is capable of. In this case, the purpose was to demonstrate the effectiveness of a feature that automatically improves performance over time, not to demonstrate a maximum IOPS capability.
David Sacks (IBM)
It's important to note that the Kaminario as priced and tested has a measly 1.1TB of storage, and the IBM Storwize just 1.5. The VSP has 11TB and IBM's DS8870 has 56. Once you calculate $/IOPS/TB, the VSP is more in the class of the DS8870 than the Storwize.
It's a little silly of me, I know, but I expect an article about storage to include information about, you know, the storage.
Thanks for that!
I staggered out of the article dazed and perplexed wondering exactly that. Thank Goddesses your comment was right at the bottom. Waiting like a St. Bernard with a cask of tiger blood to replenish my Adonis DNA. You just saved what little sanity I have left. Winning.
Odd that the word "Latency" wasn't even on the page, let alone in the article too. I could build somthing with massive IOPS if I were allowed a second of latency to return them all but a SAN with truly useful performance needs to respond before making the CPU wait too long (Violin marketing explans this well - I've just been brainwashed by them...). Given that a Modern CPU socket can do 48 BILLION things a second, a SAN doing 1 million things a second isn't all that impressive and seems a bit wasteful.
As Steve said, if you're storing a million things a second, 1TB won't last long unless you're overwriting the same data constantly. If you are, then that flash won't last long...
Yes, it is really all about latency for most workloads. I am not sure why they keep listing I/O as the performance benchmark.
Violin is alright, but expensive. TMS, now IBM FlashSystem, is the design that Violin is based on and it is less costly. IBM FS and Violin are the only true Flash arrays on the market. The rest are just traditional arrays with SSD, swapping out the spinning disk with semi-conductor on a drive and not touching the rest of the architecture.
any special tweaks?
I find it interesting that just by adding flash they more than doubled their IOPS on the platform(vs their previous SPC-1 results at least). I haven't had time to look at the details yet.
That second place comes at a pretty high cost as one would expect...interesting in any case though, nice to see more results!
"short stroking" SSD
Was just looking at the docs and this HDS system has a protected application utilization of only 53%, which is just terrible for an all-SSD platform. Other tested all flash SPC-1 systems range from 67%(Huawei) to 86%(HDS HUS - ironically enough).
I guess the flash they are using is really slow?(relative to the SLC the other platforms are using).
Missing Some Major Emerging Players
Personally I would be very interested to see what the Gemini from Nimbus Data @ 10TB and Infiniband connectivity would do in this test with test figure of around 1m IOPS.
Maybe Nimbus UK can help??
Re: Missing Some Major Emerging Players
I'm in agreement and quite surprised Nimbus are not involved in this benchmark. Tom sure loves a benchmark test and Gemini would put some of those quoted £ per IOPS to shame.
No zooming of chart with Nokia lumia 920..
.. While using the register app...
- Vid Hubble 'scope snaps 200,000-ton chunky crumble conundrum
- Bugger the jetpack, where's my 21st-century Psion?
- Windows 8.1 Update 1 spewed online a MONTH early – by Microsoft
- Google offers up its own Googlers in cloud channel chumship trawl
- Something for the Weekend, Sir? Why can’t I walk past Maplin without buying stuff I don’t need?