Disk Write Cache Feature and Limited RAM Usage with Windows 10 Pro - 256GB DDR4

Question

I have a Windows 10 Professional workstation I use to simulate material flow in impressions. The FEA software I use creates large 50-100GB database files for each simulation I run. Copying these files to spinning media for storage seems to not take advantage of the amount of RAM this system has while transferring, starting off quick for a second or two, and then dropping to the two RAID0 7200RPM disk's native transfer speed of 300 MB/s (171-342 seconds). The files are coming from a software RAID0 of two 600GB partitions on two 1.2TB Intel 750 PCIe NVMe SSDs, so read performance is not the issue. The system is on a 2200VA UPS with an EBM, and backed up nightly to our storage server, so no worries about data loss.

What I am wondering is:

If I can tweak Windows 10's cache settings to read the entirety of the 50-100GB files into RAM at the 4GB/s (12-25 seconds) the two Intel 750s are capable of, and then write them to disk transparently. I am under the impression that the built-in Windows feature "disk caching" is capable of this, but some default cache-size setting in Windows is limiting the cache to what looks like ~5GB (hence the small burst of speed at the start). I don't think this blip comes from the miserly 128MB cache on the destination drives, as "Modified" Physical Memory usage does go up by ~5GB in that first second or so of transfer. That 5GB can be seen being written to disk after the transfer dialog box disappears. RAM usage decreases in hand with the speed of the two 7200RPM drives in RAID0. Once completed, you can see the disk activity go to zero and the RAM utilization return to normal. This tells me that disk caching is at least working, just limited to 5GB at most.

The system would be fine using 50-100GB of it's available RAM for this transfer, as the simulations typically only use up to maybe 80GB of RAM, and the simulations are not using that amount of RAM until the last stages of the simulation.

I have a Dell Precision T7910 workstation with the following specs:

2P Xeon E5-2687W v4
256GB LRDIMM ECC DDR4 2400MHz
Quadro M4000
x2 Intel 750 1.2TB PCIe NVMe, one boot, 600GB RAID0 Two Partitons
x2 WD Gold 8TB in software RAID0 on SAS 12 GB Ports (128MB Cache)
Eaton 5PX 2200 IRT 240V UPS
Windows 10 Pro 1703 - System is old enough to not have Windows 10 Pro for Workstations

What I have tried:

Checked/Enabled: "Enable write caching on the device." - On each Disk
Checked/Enabled: "Turn off Windows write-cache buffer flushing on the device." - On each Disk
Made sure Superfetch service is running (for whatever good that does).
Moved away from built-in hardware RAID, as there is NO cache anyway.

I have researched other topics with similar issues, and have come across another older thread mentioning a "CacheSet" tool:

How to increase the disk cache of Windows 7

Would this be applicable to my use-case or should I keep looking?

Is my understanding of how disk caching on a Windows platform works correct, or does it operate differently than I anticipated? I am just looking for write-caching to main memory, using maybe up to 100GB of RAM, nothing else.

Thank you for your help! Any suggestions are welcome.

EDIT: Running that cacheset.exe software as ADMIN reports the "Peak size" of "663732 KB" which seems too small (648MB?). Just not sure I want to commit changing this setting and potentially messing up this in-production system. The limit I keep running into is right around 5GB.

DOUBLE EDIT: I revised the apparent GBs that seem to be actually cached. The key was looking at "Modified Physical Memory" and seeing the ~5GB "cap" at the start of the transfer. Still looking to increase this to something like 100GB.

Thank you again!

Solution

A larger cache won't help, not with any standard file copy app.

No matter how large the Windows file cache is, no sane file copy tool will close the input files for the copy task (allowing you to delete them) before all the data has been written to the destination. That is true even if the entire input data happens to have been read into the cache.

The reason is that the data in the cache is not safe - it could disappear at any time. Any RAM used by the cache that has not been modified since it was read from disk is considered discardable by the memory manager. That is, if something else needs the RAM, RAM used by the cache can be grabbed from the cache and "repurposed", ie given to the something else, at any moment. Of course it can - after all, there isn't supposed to be any data anywhere that's only in the cache. (Unless it's been modified since being read from disk. In that case it is automatically queued for writeback, which will happen within four seconds; it can't be repurposed until the writeback is complete.)

So with a standard copy program, you're going to have to wait for the spinning disks to write the data before you can delete the source files - regardless of how much is buffered in a cache.

Note that a copy program (or any other app) cannot even find out whether something is in the cache. There's no interface for that. Even if there was, the information would be considered "stale" the moment it was retrieved, even before the app looked at it. The file cache is supposed to work automatically, transparent to apps, and part of transparency is that there are very few controls on it.

Think of it this way: You need a safe intermediate copy of your results - something functionally equivalent to another copy of your source files in a different directory, maybe even a separate drive letter, before you can safely delete the originals. The Windows file cache will never give you that. Even if Windows does decide to slurp the original files in their entirety into the file cache (which is very unlikely), the file cache does not give you that.

(You may be wondering "so what good is it anyway?" The main goal of the file cache is to make repeated access to numerous small files (and file system metadata) faster. And it does that quite well.)

SuperFetch

TL,DR version: SuperFetch doesn't give you that either.

You are correct in your expressed doubt about SuperFetch. While SuperFetch is effective at what it tries to do, it won't help this case. What SuperFetch does is to keep track of files that are frequently accessed, say on every boot, and try to read them into RAM in advance of need.

This distinction is important if you want to understand Windows caching in general. The Windows file cache (which is what I've described in the previous section) is reactive, meaning that it never caches anything until a program has actually tried to read it. It has been in the Windows NT family since first release (NT 3.1).

SuperFetch is a separate mechanism, originally added with Vista. It is proactive, trying to pre-fetch things that have been observed to have been accessed often in the past.

SuperFetch manages its RAM separately from the Windows file cache. SuperFetch uses "lower-priority" pages on the Windows standby page list - and funny thing, it just leaves them on that list, so they remain part of "available" RAM. (So all other things being equal, you won't notice a difference in "Available" RAM with our without SuperFetch enabled, but you will notice a difference in the amount of reported "Cache".) The RAM assigned to the file cache is in a working set and therefore takes a little longer to repurpose, if that need arises. Because of this design the SuperFetch cache is even more quickly "discardable" than the Windows file cache, so it does not give you a safe temporary copy of anything, any more than the file cache does.

So what can you do?

To solve your problem, I would be looking at dedicated hardware. Maybe get a single not-so-expensive-or-fast SSD of say 120 GB and copy the data from your SSD array to that, then copy it from there to the hard drive. This would, alas, mean that you can't start writing to the hard drives until all of your data has been copied to the temporary, so this will take longer than what you're doing now. But the source data will be freed sooner.

A simpler idea would be to get more hard drives and put them in a larger stripe set to increase write throughput.

A dedicated copy program, one that knows the structure of your data, might help.

btw, I hope your FEA software is using mapped file access when creating the data set - it's far faster than traditional read/write calls.

Comments

[Thalys]

this looks... fun. If you don't end up getting an answer, and don't mind keeping an eye on the question for me, I might be willing to get a bounty on this

[djsmiley2kStaysInside]

I don't have an 'answer' as I can't test, but I wonder if some of the settings (the client ones) in msdn.microsoft.com/en-us/library/windows/hardware/… would help.

[tvdo]

You can use third-party tools like PrimoCache, if the built-in caching isn't enough for you. I'd suggest that at >100GB RAM you're well into specialised areas.

[GHTurbines]

Thank you Bob. PrimoCache looks like it does what I need, but it seems like this can be fixed by changing a silly limitation hard-coded somewhere in the registry (thank you djsmiley2k, looking now). Not sure I would completely trust inserting another 3rd party layer to "intercept I/O" without testing it thoroughly on a non-production system first. I will see if we can test it out on a lesser 1P Precision with another Intel 750 and spinning rust. This T7910 is our dedicated simulation box. Thank you again!

[GHTurbines]

I will keep an eye on this question for a while, Journeyman Geek. Thank you! As it stands, as long as we stay ahead of transferring off of the SSD before it becomes full (transferring during simulations), we should not have to wait for a transfer to complete to start a new simulation. 1TB of SSD is done with in less than 1 hour if we are busy, or working on a complex impression.

[AdmiralFreebee]

You say that the risk of data loss is mitigated by the use of a UPS but did you also consider the possibility of a sudden reboot or BSOD? By having a huge cache in RAM you're taking a higher risk of losing data during a copy operation. Perhaps it's an acceptable risk but I wanted to point it out nonetheless.

[GHTurbines]

I can see your point with a cut+paste operation. A copy+paste and delete would require waiting for the transfer regardless.At this point, it may be best to get faster storage or go 10 gigabit to a network location capable of 1 GB/s. I can see Windows 10 being unstable even with decent hardware.

[Damon]

Why not use a RAM drive as a buffer?

[Jamie Hanrahan]

Stuff copied into a RAM drive isn't safe. Also, an SSD of around 250GB will cost a heck of a lot less than a similar amount of RAM.

[Damon]

A buffer is inherently a temporary SECOND location where long term "safety" is not a concern. Having available RAM for the files is part of the premise of the question, he has nothing to purchase for a RAM buffer.

[Daniel B]

Well technically I think he doesn’t really want to to be cached. He wants it to be buffered. That’s safe, of course, but still won’t provide any advantage whatsoever.

[Jamie Hanrahan]

You are correct, but my strong impression is that he was asking about getting the Windows file cache to do the job. Hence my discourse about why it isn't suitable.

[user2284570]

Compared this to Linux, where I can get a file copy to finish instantly though increase the write behind cache to take all the availaible ram, and act as if the file copy is already done (relative to file path). And where starting syncing with the kernel cache can be done only manually (if the option to do it periodically is disabled).

[apraetor]

That's a very good point, Jamie. OP's goal seems unclear, beyond wanting to literally read everything into a ram buffer asap. And as you explained, there's no point -- the underlying write operation will still bottleneck at the write to the array. Doesn't matter whether it's buffered first or not. BUT OP does have a 12GB SAS card... there's no way they are saturating those links with only 2 disks in the RAID0. Like that KMFDM song, more and faster!