Performance problems in L1 - CrystalDisk behavior Topic is solved

[CescCerdanya]

Hi all. This is my first post. My English is not good, so I help a translator. I'm running PrimoCache 3.0.2 in its test version.

I expose: My computer is an i7-4790 with 16GB DDR3 1600. 1 SSD Samsung 860 EVO 500GB for the OS (C:\). 1 HDD Seagate 2TB for data (D:\). 1 Kingston A400 SSD for HDD cache.

I created an L1 cache for C:\ of the following features: 4096 MB of RAM, read and write, Block size 8KB, idle-flush defer-write - free cache on written - flush on sleep, prefetch last cache - start at windows boot (This last one did not work until you did not execute the command 'rxpcc set BootFetch 1').

I have also created a read-only L2 cache for D:\ in Kingston SSD that also generates some doubts that I will not try in this post.

When installing PrimoCache it had only 8GB of RAM and the synthetic tests of CrystalDiskMark were:



Then install 8GB more identical RAM to run on Dual Channel and the results were:



Records in Seq Q32T1 almost doubled, and the others remained stable. I was surprised by the increase of the sequential test but I attributed it to the Dual Channel (128 bits).

My surprise has been that a week later, and without having modified anything, the readings are now:



The records of the sequential test have been halfway through the first two tests but all the others have dropped by half or more. I have run the test several times. After rebooting, deleting the PrimoCache task and re-creating it, and always the same values. I have reviewed the UEFI values ​​of the memory, the same ones. I have passed the AIDA64 memory tests, same results. What has changed? What am I doing wrong?

Thanks in advance for the help.

[cichy45]

That might be due to Spectre and Meltdown (and other Intel vulnerabilities) patches that installed in some update. They have huge impact especially on I/O. So that is why your 4KB random I/O much lower than sequential performance on bigger chunks of data (so less I/O). That is one possible explanation.

[CescCerdanya]

Maybe, but it's very unlikely. The first image is from October 12, the second from December 2 and the third is from today, December 19. I do not know that any important vulnerability has been published on these dates.

Thanks for answering.

P.S. In addition, the results of the AIDA64 memory tests would also be affected, and this is not the case.

[CescCerdanya]

Solved. For a while I thought I was crazy.

It is a bug, in my opinion, of CrystalDiskMark (P.S. Or a restriction to processes with low privileges for the Spectre / Meltdown patches). If I executed it from a shortcut on the desktop, the records were slow. If I executed it from the file manager I use, Altap Salamander, the records were fast.

Finally I realized that Altap Salamander executed it with Administrator permissions. If I run Crystal with administrator permissions, fast records, if not, slow records.

In the image the two side results with two instances of Crystal, the one on the left executed without Administrator permissions and the one on the right with permissions.



Someone had appreciated something similar?

P.S. I see that other people have found the same
https://linustechtips.com/main/topic/95 ... rk-speeds/

[Jaga]

I generally tend to stick with AnvilPro when doing comparison benchmark testing on drives. CrystalDiskMark is okay to get a "general feel for what's going on", but not when you change things up or want to do a more detailed analysis.

[Support]

Interesting! We also observed such behaviors from Crystaldiskmark.

[CescCerdanya]

The problem is that CrystalDisk is allowed to run without Administrator permissions. Both AnvilPro and AS SSD ask for permissions (if UAC activated) and they are not allowed to run without privileges, and therefore their results are always consistent.

On the other hand, if the Spectre/Meltdown patches slow down the system only if processes are executed without Administrator permissions, as it is deduced from the image posted in the post to which I mentioned...



...perhaps the behavior of Crystal is the most honest.

[Support]

@CescCerdanya, thank you for sharing the information!

[CescCerdanya]

You're welcome. I want to point out that the most disturbing aspect of this behavior is the fact that the patches do not affect if one executes a process with Administrator privileges.

Does this mean that processes executed with privileges are vulnerable to Spectre/Meltdown? How many processes with privileges do you have running on your machines? Is this a question for Microsoft?

good morning to everybody

[cichy45]

I guess that no process except for system core (and a few other related) are running as root/admin. I think that even some core components do not use admin privileges. Windows always execute them as user except when you force to execute as admin. Maybe PrimoCache process/service is also running as user and therefore with lower-than-possible performance? Can @support provide us with such info? Maybe we can force it to run at admin level.