Hardware for my new server

[rungekutta]

RAID5 or RAIDZ write performance will be that of the slowest disk (as all disks have to finish writing before the write is complete), plus overhead in the controller or software. Laughable or not, that's how it works and it's down to pure logic and physics. You can get more performance out of it (as above example) by striping vdevs and in some cases by using an SSD SLOG.

 

[Stux]

RAID5 or RAIDZ write performance will be that of the slowest disk (as all disks have to finish writing before the write is complete), plus overhead in the controller or software. Laughable or not, that's how it works and it's down to pure logic and physics. You can get more performance out of it (as above example) by striping vdevs and in some cases by using an SSD SLOG.

But, the slowest disk is writing 1/(n-p) part of the original data and thus will be faster than a single drive writing all the data?

My tests (admittedly simple) have verified a r/w performance gain from raidZ.

 

[rungekutta]

But, the slowest disk is writing 1/(n-p) part of the original data and thus will be faster than a single drive writing all the data?
My tests (admittedly simple) have verified a r/w performance gain from raidZ.

Yeah maybe so. I guess it will depend on file size also. Many small writes will typically be significantly slower than large continuous ones due to seek times (latency) on the drives themselves. And to the extent latency varies (depending on where the head was positioned just before etc), the complete operation will again be punished by the maximum of the seek time of the drives involved i.e. the larger number of drives, the closer to the worst outcome every time. With larger continuous reads/writes this should matter less, as the seek time is relatively small relative to the read/write operation itself (until the drive starts to get heavily fragmented?).

Some benchmarks here regarding drives in different setups (for what they're worth): https://calomel.org/zfs_raid_speed_capacity.html

 

[Helios]

The performance depends on the specific kind of write. Writing a full stripe is no problem, the throughput is slowest_drive * drives_in_array - parity_computation_in_memory. Writing only part of a stripe, such as when updating a file or when copying small files, is more problematic. The system needs to read the full stripe, check it, update whatever block are necessary, compute parity, and write back. The worst case performance for raidzn is when single bytes on random locations have to be updated.
Those are the benchmarks I talked about. Raidz2 over six disks performed at 429MB/s, while a single disk performed at 108MB/s. I really don't know how that's possible. I can only guess that ZFS is taking full advantage of memory and CPU resources, which hardware RAID controllers can't do.

I just got the drives, and I'm fairly impressed with them. Around 140 MiB/s of sequential access. It actually beat the Black on random 4K access, although the Black is 3/4 full, so that skews the results against it because the benchmark program can only move the head around the tracks closest to the spindle, where the HDD performs worst.
I haven't tested in ZFS yet. I didn't feel like completely taking my computer apart just to try them out. I should have everything set up by next week. I'll try to post some numbers then.

 

[Mryusef1]

has anyone used the Asus z9pr-D12 with freenas. I just picked one up with two CPUs but can't get any video out of it but I can log in to it through ssh. Any ideas?

 

[David Henrickson]

has anyone used the Asus z9pr-D12 with freenas. I just picked one up with two CPUs but can't get any video out of it but I can log in to it through ssh. Any ideas?

I highly recommend starting your own thread about this. I wouldn't use a 4 year old dead thread to post something that doesn't deal with the persons initial questions. haha...just my thoughts. I would def start another one. People are pretty quick to respond, and that way you can get direct answers to your question related to your hardware. Thats all :)

 

[Mryusef1]

OK will do