PrimoCache vs SuperCache 4K file Read caching?

[NessPJ]

Hey guys,

I recently started my trial with your PrimoCache software. So far i think the software looks good and it seems to be working very nice.
I see a lot of people on these forums using Crystaldiskmark for posting their results. I always use AS SSD for benchmarking as it seems to give more accurate results as to what is happening (where CDM tends to color most numbers much more positive...).

I have some past experience with using SuperCache as a RAM caching tool, and i noticed that my 4K random read speeds seem to be much much lower then when i was using SuperCache before.
With SuperCache i was hitting around 300 MB/sec random 4K read speeds. While, with PrimoCache i seem to be around 34 MB/sec random 4K read speeds at most!
Any idea where this could relate too?
I have been using the same type of Cache specs on both pieces of software: 2048MB or 4096MB Cache size. 4K stripe size.

Here's some information of the hardware i am using:
Intel DZ87KLT-75K
Intel Core i7 Hasswell 4770K
2 x 4GB G.Skill RipjawsX 1600MHz DDR3
4 x Sandisk Extreme II 120GB SSD's in RAID0.

Any idea, why PrimoCache seems to be hitting significantly lower 4K Read speeds?
(I ask, because these are really the ones that matter).

Thanks in advance.

[Support]

Do you use same test data size? Your PrimoCache settings? Thanks.

[NessPJ]

Hi there,

Yes, the same test data size's are always used to ensure an objective test.

SuperCache 5 (read-only mode):
4K random read = 308 MB/s
4K random write = 79 MB/s

PrimoCache (read-only mode):
4K random read = 34 MB/s
4K random write = 58 MB/s

[CrypEd]

Good 4K-rates here using one single drive while CPU is downclocked.

good4k.jpg (96.7 KiB) Viewed 13196 times

1. We asked for your Primo-Cache-Settings... this means: LRU or LFU-R, which cache-size, which blocksize?
2. I like to know what is the test-size in specific, not only if it is the same: What is it please?

(3. SInce you not specified your HW-Raid-Controllers hardware details I assume that you use an onboard Firmware-Raid-Chip. You need to know that 4k-R/W use some decent CPU and in a FW-Raid-Setup uses additional CPU-usage, so it's probably not the best setup to achieve high-4k-rates.)

True thing seems to be that Supercache shows higher rates for you.... but without reviewing your PrimoCache setup this means nothing. For example: Your result is quite good for 512byte-blocksize, 128Mb cachesize, configured with LFU-R.

While I think it's funny if someone buys 4x SSD to maintain a FW-Raid-Controller while you can buy dozens of PCIe-Cards with SSD chipy maintained internally as a raid0 by design.

[NessPJ]

Hello there,

Thanks for your reply.
You seem to be posting some very impressive speeds (even though CDM does not bench with uncompressable data...yet this usually does not color the results for more then 10-15%).

Basically all of the information you asked for was in my OP (Cache size = 4096MB, Stripe size = 4K).
I did not mention the algorythm (my bad), i'm using LFU-R.

Yes, i am using the on-board RAID controller (Intel's first using 6 x SATA3 interface....which would imply a lot of overhead before the controller would get saturated...sadly this was not the case).
I know 4K I/O are very CPU intensive, but i hope we can both agree that for the current mainstream market it doesn't get much better then a 4770K.
So even if this setup would not be most ideal for achieving the highest 4K rates, i still get around 350 MB/sec 4K reads with supercache (and from what i have seen/read so far... PrimoCache can often outperform SuperCache, which is where it shows its diversity...)

Also there is a definite reason i did not buy a PCI-e SSD card. Yes, i bought 4 SSD's and this still saves me quite some money compared to a decent PCI-e card (i got these for little over 400 EUR. total getting 480GB storage as a result).
Next to that, all external controllers found on PCI-e cards add to the systems boot time (some as much as 18-20 seconds!)....which was a goal in itself to try and achieve short boot times.

I hope this clarifies a few things on my side.

[InquiringMind]

NessPJ,

Could you specify the block size you are using in PrimoCache? If using the default (4KB) that will likely explain your performance figures - if so consider increasing it. I found 64KB to be the best, results posted in the Benchmarking PrimoCache thread.

Uncached results may be of interest too - your setup should be delivering better performance than it seems too.

While I think it's funny if someone buys 4x SSD to maintain a FW-Raid-Controller while you can buy dozens of PCIe-Cards with SSD chipy maintained internally as a raid0 by design.

PCI-E SSD cards can offer better bandwidth but currently cost (much) more per GB, don't work in all motherboards with SLI/Crossfire (usually due to motherboard issues limiting use of a third PCI-E x16 slot - affects hardware RAID controllers too) and can't be upgraded.

[Support]

@NessPJ,

Is the test size 1GB by default or a value less than cache size? Can you try a test immediately after boot completes so that cache is empty for the testing? Thanks.

[CrypEd]

Im just saying that using FW-Raid0, or even HW-Raid, whatever it is based on a SSD-internal-controller (PCIe-Cards) or Hardware-Card-Controllers with SSD-Driver at all, is not what increases 4K. While 4K-scenarios are a little less significant overall, since you barely know a real-life use case, that takes use of large amount of 4k-R/W.

Im just saying single drive SSD will outperform any raid0-SSD array in 4K-disciple, while on the other hand that disciple itself isn't of such a high significance.

PCIe cards are great, because they maintain seperated SATA3-controllers single each ~600mb/s per internal raid0-port, while such raid0-controlling complety is hardcoded into the firmware, often containing a LSI bridge chip. These cards offer superior sequential performance. The good thing is, that these cards report to the BIOS/OS as a single simple drive. YOu do not need any CPU-cycles extra maintaining data on such storage card, thats another benefit.

You do not need to configure anything.

It is not true that systems with PCIe-SSD-Card added need significant longer boot-up.

It is tot true that these cards are more expensive. You could have it at nearly same price.

Whatever, I don't want to start a flameware upon that. I also agree with you that with your raid- or primocache-configuration, does not operate well. The problem is to think it's primocache's fault only because supercache does not have this issue, doesn't mean it's probably the drivers or the FW-raid-controllers fault... and maybe the supercache-devs found a workaround or something. It's not that easy, doh.

Still I say that whatever you have 40mb/s or 350mb/s 4k-R/W doesn't make any significant difference in real work circumstances in which 4k-RW usually come ahead (less amounts). If you compare 40mb/s 4k-R/W to a HDD, that for example would still be pretty impressive.

One thing is you are using LFU-R, that means it does NOT keep recent data (for example bench-data) but keeps the data you most frequently use. CHance is high your cache is full. Does the 4k-rates improve when you re-run the 4k-test a couple of times? Could you check LRU?

When asking for primo-caches "block-size", we doesn't mean how you configured your FW-raids-stripe size, in the controller settings. In case it is 4k blocksize you indeed could try to increase it a few steps and see if the result gets better.

[Bjameson]

For what it's worth, I compared Primocache and Supercache on a single-drive SSD (Samsung EVO, Asus mobo). Both 4K blocks, both 4 GB cache. No RAID. Using Atto disk benchmark.

Primocache outperformed Supercache by a factor 2. Supercache disallows write caching on SSDs and even slowed down reads compared to no cache at all. On spinning disks, Supercache performs very well. However, Primocache speeds up both reads and writes on all drives. I'd been considering buying Supercache but now it looks like a waste of money. At least on non-RAID volumes, Primocache looks like the clear winner.

@Romex: could you please give us a price indication for the final version? Will there be separate home- and enterprise versions? Thanks for any info you might be willing to give. Keep up the good work!