Fiddling with the newest PrimoCache (3.0.2) version

[BonzaiDuck]

I know I have a "rep." I started using PrimoCache in 2014, now on four of my systems. My Skylake rig uses a Sammy 960 EVO 256GB NVMe M.2 to cache an SATA SSD and an HDD.

In the old versions, you could see the result of RAM-caching a drive immediately with a benchmark like Anvil, but you couldn't see it for SSD-caching of a slower device. This was due to the fact that the cache fills up when the system is idle: a benchmark wouldn't register the improvement.

The new version 3.0 suddenly makes up for this shortcoming, and you can divide an SSD caching volume into a Read and Write portion. I'm still exploring the implications of that feature, and I'm even a little bit confused by it. However, I can now run Anvil against my Crucial MX 300 -- cached to the 960 EVO with no RAM-caching -- and it now shows how even a decent late-model SATA SSD improves in performance with the NVME M.2 cache:

The Crucial by itself would only have a sequential read speed of about 500 MB/s, so this confirms the operation of the cache. I'm still wondering about the 4K random read test, however. It doesn't seem different than what would occur for an uncached SATA SSD.

The Write results shown here require that the deferred write feature of the cache are turned on and enabled. Otherwise, you would only see the spec expectations for the MX 300.

 

[Doomguy]

Did you ever get the performance on the 4K reads to improve?

 

[BonzaiDuck]

Did you ever get the performance on the 4K reads to improve?

Only by creating a two-tiered cache using both RAM and the NVME 960 EVO:

If the EVO were replaced by an Intel 900P, the 4K results would look similar to this, but the RAM-caching would be unnecessary. Then -- imagine what the bench results would look like with the RAM-caching . . .

 

[nosirrahx]

If the EVO were replaced by an Intel 900P, the 4K results would look similar to this, but the RAM-caching would be unnecessary. Then -- imagine what the bench results would look like with the RAM-caching . . .

I set up a tiered system that stacks 12GB of primocache RAM on top of a 260GB 900P U.2 which is caching a 4TB 860 Evo.

This system has a UPS and deferred writes enabled with a 5 second delay.

Optane and Primo can absolutely be used together to accelerate different aspects of a system.

Primo wont help booting at all while Optane absolutely does. Primo on the other hand accelerates the live system more than Optane can.

These are the 900P as Optane cache with Primo disabled and then with it enabled:

 

[BonzaiDuck]

I set up a tiered system that stacks 12GB of primocache RAM on top of a 260GB 900P U.2 which is caching a 4TB 860 Evo.

This system has a UPS and deferred writes enabled with a 5 second delay.

Optane and Primo can absolutely be used together to accelerate different aspects of a system.

Primo wont help booting at all while Optane absolutely does. Primo on the other hand accelerates the live system more than Optane can.

These are the 900P as Optane cache with Primo disabled and then with it enabled:

I can't tell which 4K scores are comparable to the Anvil scores. Wait-a-minute . . . . without taking time to upload the screenie, my CrystalDiskMark scores are:

14650 / 13387
2451.7 / 1791.5
1263.3 / 766.5
1066.7 / 671.0

That's with an 8GB RAM-cache on top of a 140GB cache on a 250GB 960 EVO. The deferred write delay is the default 10 seconds. I won't dispute your conclusion, but I'm puzzled by my low 4K write scores. Or maybe that's just the strong points of your 3D Xpoint Optane showing through.

And why do your Primo scores fall down in the second and third tests over the Optane cache? But with any of those scores, I guess it's so good it doesn't much matter.

Mmm . . . let me see about changing the deferred write delay: Nope. A 5 second delay only makes the last two write scores lower by 30 to 80 points.

If the 900P had been available in January 2017, I would've bought one over the 960 EVO. But those are just the breaks, I guess . . .

Man! That 860 EVO 4TB costs a pile of change! I thought I spent a lot on a 2TB Crucial. . . .

 

[nosirrahx]

And why do your Primo scores fall down in the second and third tests over the Optane cache? But with any of those scores, I guess it's so good it doesn't much matter.

Its due to me not setting the same parameters.

In the first test I was using Q4T4 and Q8T8.

I have switched to Q1T2 and Q2T1 to get a better idea of how things scale up from the worst case scenario.

 

[BonzaiDuck]

Well . . . I don't think the 900P is going away very soon. They'll still be available in the next year's time. But with the performance I now have with the little EVO drive, I'm sure I can wait a while. I also have to think about what sort of system I'll build next. And I can also take my precious time with that as well . . .

Like XavierMace says and assuming one has freely overinvested in benchmark performance, you'd never notice the difference much beyond consumer-systems with only an NVME drive as enhancement. [I like to THINK I notice the difference, but it could be like the "placebo-effect." Any way I cut it, it's way-y-y-y fast enough.

I'm still interested in your test results, though. What did I hear about the TBW life expectations for the 900P? Maybe someone posted an indication here or in another thread. Or maybe it's just in the Intel spec sheet.

 

[nosirrahx]

I'm still interested in your test results, though. What did I hear about the TBW life expectations for the 900P? Maybe someone posted an indication here or in another thread. Or maybe it's just in the Intel spec sheet.

3dXpoint (Optane) is just about indestructible when it comes to consumer usage. I could use this PC every day for a decade and barely put a dent in its life expectancy.

https://ark.intel.com/products/1236...00P-Series-280GB-2_5in-PCIe-x4-20nm-3D-XPoint

 

[IntelUser2000]

3dXpoint (Optane) is just about indestructible when it comes to consumer usage. I could use this PC every day for a decade and barely put a dent in its life expectancy.

https://ark.intel.com/products/1236...00P-Series-280GB-2_5in-PCIe-x4-20nm-3D-XPoint

5.11PB means you need to write 1 terabyte a day for 14 years. Endurance specs are also the worst case scenario. So you are talking 4KB, QD1 random writes. You probably have a much better chance of random component failures before reaching rated write figures. All electronics have a decent chance of failing after 10 years of usage anyway.

With the 900P, it won't even brick the drive after the lifespan is reached. After 105% lifespan is reached, the 900P goes into ultra low write speed mode. That means the failure mechanism is more akin to HDDs where its more graceful and still semi-usable, unlike SSDs where it can be quite abrupt.

 

[BonzaiDuck]

Sounds like a winner. Beyond the expectations of our puny little brains . . . .

 

[nosirrahx]

Sounds like a winner. Beyond the expectations of our puny little brains . . . .

Its on the right track for sure. The problem with Optane currently is the price, form factor and capacity. The "good" drives are U.2 and add in cards. The M.2 Optane drives have low capacity and are way slower than the 900P.

Currently I would suggest Optane only to enthusiasts and even then only the 900P drives. A U.2 900P combined with that massive SATA SSD is the only way to get huge capacity and impressive speed, its just not economical especially in the era of insane RAM prices and insane GFX card prices.

It looks like Samsung is about to close the gap with some new technology. It wont be as fast as Optane but its close and I am willing to bet that the capacity, price and form factor will all be better than Optane.

When PCIe gen 4 makes it way to motherboards we likely will see an Intel VS. Samsung performance war when it comes to storage, its going to be very interesting.

 

[Brahmzy]

Question. I’ve messed with Fancycache way back.
I’ve got 2 PCs in particular, both with 280GB 900p’s for the OS. All of the apps on both live just on the 900p’s. In this case, PrimoCache isn’t going to do much for me, eh?
They’re both already ridiculously fast, but always looking for faster.

In this instance, both machines could give up about 8GB of their 16GB of memory, but how much can Primocahe L1 do/improve, over what W10 is already doing?

 

[nosirrahx]

Question. I’ve messed with Fancycache way back.
I’ve got 2PCs, both with 280GB 900p’s for the OS. All of the apps on both live just on the 900p’s. In this case, PrimoCache isn’t going to do much for me, eh?
They’re both already ridiculously fast, but always looking for faster.

The system I use the 900P and primo cache together on uses 32 gigs DDR4 3833 (could not get higher with 4 sticks) and a 8700K delidded @ 4.8GHZ.

Under these conditions 12GB of RAM cache actually makes a difference in some cases. Re-launch of apps is close to instantaneous.

That said the gap between a HDD and the 900P is so vast that any increase over that jump feels far less significant.

 

[BonzaiDuck]

Personally, if I couldn't choose a 500GB 900P over a 280GB 900P for a boot-system volume, I'd be more inclined to spend my money on a larger SATA SSD and use the 280GB 900P to cache it. In fact, depending on personal usage pattern, I might not even bother with the SATA SSD and might just spend $50 to $100 on a good HDD to cache the same way. If I could put my OS on a 900P drive >= 500GB and I had surplus RAM or felt that I could spare 4GB or more, I'd still cache the 900P to RAM.

 

[flexy]

I am giving Primocache a try on my media HTPC/torrent machine. So far I like it, but I am not sure whether it's really worth it.

The issue with my torrent box is that it has slow mechanical HDs.

The box runs Win 10 x64, and since it's really only used for watching movies and torrenting, it has 8GB of memory, which for these things is totally sufficient.

I set Primocache to use 3GB of memory for L1 cache, which I set w/ the slider to 90% read, 10% write.
I only want to possibly "speed" up seeding torrents, so I am merely interested in read caching, *IF* this even has only benefit.

The OS partition of the box usually shows me 64% cache hit rate.
The F: partition (which is on the same drive) holding all the torrents most of the time only has 6-7% cache hit rate.

Obviously, most of the torrent files are several gigs, with many 2-3 gigs and others often much larger.

Since I am rarely seeing more than 7% cache hit rate, I wonder whether Primocache in my situation even makes sense.

Any advice?

 

[nosirrahx]

I set Primocache to use 3GB of memory for L1 cache, which I set w/ the slider to 90% read, 10% write.
I only want to possibly "speed" up seeding torrents, so I am merely interested in read caching, *IF* this even has only benefit.

If your internet is fast enough that uploading files is hindered by mechanical HDDs then I am crazy jealous.

For large file cache you would want to use a SSD, not RAM (also something Primo handles).

 

[flexy]

I really only installed Primo since a lot of memory on that box is otherwise not used, so I figured why not, and to see whether it might reduce access times to torrents. (I did not install it out of a need, nor can I actually say whether it benefits anything on that box)

But now you made me curious since I actually never considered that the speed of some old HD itself might limiting my upload rate. My internet is 500/500. It's not super-duper fast, but still decent. The HD is an old Hitachi Deskstar, which should yield on avg 60MB/s.