would setting sync off on a smb share improve it's speed?

[Yorick]

thank you. what might be the reason for those misses? honestly i might be asking a stupid question but misses sound to me like something that shouldn't happen. is the system out of sync or something?

Not enough RAM to keep all the metadata in ARC. Whether having more RAM for metadata, or a small metadata-only L2ARC will help, is a really good question. This is troubleshooting - educated guessing - at its finest.

does this mean the nic is good?

We literally don't know. If you can beg borrow steal an Intel NIC somewhere, that'd be a good test. At least it'll tell you whether the NIC was part of the issue.

so hope i can find ssd that small (or should i resize it myself?).

Well, the Intel 320 40GB is 15 bucks on eBay, and, I get the issue with "Israel". Here's the formula for L2ARC RAM usage:
(L2ARC size in kilobytes) / (typical recordsize -- or volblocksize -- in kilobytes) * 70 bytes = ARC header size in RAM.

However, you are only going to use it for metadata. I see something like 120MB of metadata, so if that's kept in L2ARC, we are looking at what, under 2MB of RAM usage? Okay, you are fine. It's an easy test. I don't, really, expect that to have been the issue - and, it's an easy test.

Did you verify that the dataset or share isn't set to force sync write?

The drive is CMR, so your issue is not with DM-SMR, that at least we can say for certain.

 

[tomerg]

Not enough RAM to keep all the metadata in ARC. Whether having more RAM for metadata, or a small metadata-only L2ARC will help, is a really good question. This is troubleshooting - educated guessing - at its finest.



We literally don't know. If you can beg borrow steal an Intel NIC somewhere, that'd be a good test. At least it'll tell you whether the NIC was part of the issue.



Well, the Intel 320 40GB is 15 bucks on eBay, and, I get the issue with "Israel". Here's the formula for L2ARC RAM usage:
(L2ARC size in kilobytes) / (typical recordsize -- or volblocksize -- in kilobytes) * 70 bytes = ARC header size in RAM.

However, you are only going to use it for metadata. I see something like 120MB of metadata, so if that's kept in L2ARC, we are looking at what, under 2MB of RAM usage? Okay, you are fine. It's an easy test. I don't, really, expect that to have been the issue - and, it's an easy test.

Did you verify that the dataset or share isn't set to force sync write?

The drive is CMR, so your issue is not with DM-SMR, that at least we can say for certain.

been doing some reading today and i understand now that hits/misses refer to fetching from cache. also tried to understand what metadata is but i certainly have to read more. so what i don't get is how come there was still a lot of available memory (7GB) and still i ran out of space for the metadata. maybe i got to allocate more size for the arc? can i know how big is that metadata?
also in the choice between l2arc and memory - is it just a matter of price? (i assume ram is much faster the ssd).
dataset sync was to inherit so i understand it is ok.