L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

FAQ, getting help, user experience about PrimoCache
Post Reply
Mecky
Level 2
Level 2
Posts: 6
Joined: Fri Jun 11, 2021 11:22 am

L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Mecky »

Hello,

as a new (trial) User I might get something wrong, but I am really confused:

Preface:
A VM (on proper Hardware & ESXi) has some unoptimized (closed source) code I cannot alter.
So the idea is to improve (un)snappiness with Optane (900p) with PCIe passthrough.
The default SSD (Raid 1, 2 SATA drives) shall be accelerated (mostly in terms of latency, hence Optane).

Set 900p as L2, Enable write defer 1s, Write mode: Average
(should be "reasonable" in terms of data integrity, I think(?))
CrystalDiskMark "Real World Performance + Mix" now shows numbers for main drive as if it was Optane, that's what I've expexted / hoped!

Now to my quesion:
Adding 4096MB of L1 with 100% read now also accelerates writes (7µs RND4K!)
That I didn't expect, and don't wanna have (in terms of data integrity?)

Also, the "Total Write L1" increases on write operations?

I first thought: Yeah, that's supposed to happen, as the read-cache needs to be filled, but then again, why does the write-speed latency drops below capabilities of Optane?

Is PrimoCache too "smart" for CrystalDiskMark, or what am I missing?

Grateful for any hint in the right direction (or the correct search terms)!

Yours, Mecky
janusz521
Level 5
Level 5
Posts: 51
Joined: Wed Aug 26, 2020 6:11 pm

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by janusz521 »

Yes, PrimoCache is fooling disk benchmarking software. You are writing and reading data in PrC cache memory while the benchmark assumes it is exchanging with the disk. You will get very fast writes when you have deferred write enabled. I got similar results with Passmark Performance Test.
Mecky
Level 2
Level 2
Posts: 6
Joined: Fri Jun 11, 2021 11:22 am

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Mecky »

Sure, but it's faster than optane, so I assume RAM is involved (which I set to 100% read!)
That's what I don't get.
Mecky
Level 2
Level 2
Posts: 6
Joined: Fri Jun 11, 2021 11:22 am

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Mecky »

Still have not figured out, how a 100% read L1 can enhance write speeds on CrystalDiskMark... Am I alone with this observation?
User avatar
Support
Support Team
Support Team
Posts: 3623
Joined: Sun Dec 21, 2008 2:42 am

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Support »

Could you open your PrimoCache, choose/highlight the target volume item in the main dialog, and then make a screenshot of the PrimoCache main dialog and sent it to us? We'd like to see how the cache is configured and used. Thanks.
Mecky
Level 2
Level 2
Posts: 6
Joined: Fri Jun 11, 2021 11:22 am

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Mecky »

Yeah, sure!
Only going by the numbers of access time (higher than to be expected from Optane because of virtualization-overhead...)
As soon as L1 (100% read) is activated, write is also absurd fast:
Primocache.PNG
Primocache.PNG (565.59 KiB) Viewed 1069 times
For reference without L1, otherwise same config:
PrimocacheL2only.PNG
PrimocacheL2only.PNG (175.28 KiB) Viewed 1069 times
As I learned, data integrity of L2 isn't given (no intend-log), so there might be no benefit from this setup? Nevertheless unexpected behavior, right?
User avatar
Support
Support Team
Support Team
Posts: 3623
Joined: Sun Dec 21, 2008 2:42 am

Re: L2 (Optane) R/W L1 100%R, Writes are too fast for RO!

Post by Support »

We did a quick test following your configuration but don't see any problems. Please note that Windows has its built-in write caching with a latency of 2 seconds, maybe because of this?
You may also reset statistics data by right-click on the statistics panel and then choose "reset" on the popped-up menu. Thus you can clearly check if the value of "Total Write (L1)" is 0 after some writings.
Mecky wrote: Tue Jun 15, 2021 6:07 am data integrity of L2 isn't given (no intend-log), so there might be no benefit from this setup?

exactly, no benefit currently.
Post Reply