From: Dirk Steinberg [mailto:[email protected]] Sent: 星期三, 三月 09, 2016 0:06 To: [email protected] Subject: Re: [smartos-discuss] Samsung 950 Pro on SmartOS?
Am 07.03.2016 um 06:48 schrieb Fred Liu <[email protected]>: 2016-03-05 21:19 GMT+08:00 Dirk Steinberg <[email protected]>: > Apart from that: is NVMe support in SmartOS considered stable? There is driver support for it in the system. I have not heard many reports positively or negatively about it. [Fred]: I am testing some Intel P3600 NVMe SSD. In normal workload, they just work . But in burning mode like continuous scrubbings, I have got lots of checksum errors. And I tested the same scrubbings under Linux, no checksum errors were found. Fred Hi Fred, do you attribute these errors to the SmartOS NVMe driver? Sounds like it since you are saying that the same SSD works under Linux. So have you given up on NVMe on SmartOS? [Fred]: I personally attribute those errors to the immaturity of the NVMe driver in Illumos. But that is not so severe based on the fact that there are no substantial data loss with those checksum errors. I used to get some some kernel panics under "too may check sum erros" and after disabling "sha512|skeun|edonr" checksum algorithm, the server has been running well for more than two weeks. It looks like "sha512|skeun|edonr" checksum algorithm still has some glitches. Thanks. Fred Sorry, I am not quite sure what you are trying to tell me. I am running the sha512 hash on a regular disk pool without problems. I have also run a number of scrubs. I assume the the sha512 hash is working correctly. If I had a ton of checksum errors from ZFS I would be VERY concerned. Just turning off the checksum may not be the solution… I do not think that a bug in sha512 is generating these. Also, what do you mean by „no substantial data loss“? Either there is data loss or there is no data loss. My policy here is that even a single bit of data loss is not acceptable. This is why ZFS goes to all the effort of using strong checksums and redundancy and scrubbing and so on… If you are willing to accept „some nonsubstantial data loss“ and turn of checksums so as not to be bothered by all these checksum errors there is not much point in using ZFS in the first place! Am I understanding something incorrectly? [Fred]: "NVMe support" and "[sha512|skein|edonr] hash algorithm" are the major spot-lights of Illumos in 2015. I am running 20160218T022556Z now. Disabling "sha512|skein|edonr" doesn't mean setting "checksum=off". In default("checksum=on"), zfs automatically selects an appropriate algorithm if these features("[sha512|skein|edonr]") are enabled. I met random server reboots(triggered by zfs deadman) and panics. But I could not get the nitty-gritty reason from core dump. And we do have very stable running of release of 2014. To reason by exclusive method, I recreated the zpool and set "checksum=sha256". And till now the server has been running well at least for two weeks. This operation does not intend to remove the *annoying* checksum errors. As for "no substantial data loss", that means "zpool status" always show "errors: No known data errors" even with "too many checksum errors". Thanks. Fred ------------------------------------------- smartos-discuss Archives: https://www.listbox.com/member/archive/184463/=now RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00 Modify Your Subscription: https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb Powered by Listbox: http://www.listbox.com
