Re: [exim] Advice: NFS, hardware, SATA vs SAS etc

Góra strony
Delete this message
Reply to this message
Autor: Jasen Betts
Data:  
Dla: exim-users
Temat: Re: [exim] Advice: NFS, hardware, SATA vs SAS etc
On 2019-12-04, venbian via Exim-users <exim-users@???> wrote:
> Hello gurus,
>
> As 2020 nears I wanted to ask opinions about the current state of
> hardware requirements for a small business email platform. $dayjob
> asked me to enhance our existing platform to improve performance and
> add redundancy.


Q1: how much of each?

> Main questions (TLDR):
>
> Can NFS handle heavy IMAP, LDA, HTTP workload?


yes: this is not microsoft - there are no software performance limits.
But using NFS for server disk will double your network traffic (or
worse when doing server side searches)

> Is direct attached SATA III 6Gb/s SSD in RAID 1 sufficient or is SAS needed?


see Q1

> Is gigabit ethernet the bottleneck in any case? I can upgrade to a
> 10 gigabit local network if advisable.


> Can NFS peacefully co-exist with other mail system workloads without
> resource contention such as SQL DB or spamassassin or redis, etc?


if your database is on direct attached disk then yes.

> More detail:
>
> I had bad experiences with NFS approx 15yrs ago where IMAP load
> saturated controller link (yes, noatime was used on the mount) and was
> unusable. But in 2020 is it time to give NFS another look?


why would you want to use NFS? I would pick E-SATA or USB3
direct-attach over NFS, and internal direct-attach over that.

> Current mail storage setup uses local attached large SATA SSD and
>does well, but it directly hosts HTTP, LDA, IMAP and Submission which
>could all be faster and it only does nightly backups. Adding SAN is
>probably out of $dayjob pricerange and SAS is borderline. Power
>consumption is also a factor so instead of a dedicated file server I
>thought it would make more sense to build a big server with direct
>attached fast SATA SSD in mirrored RAID that also has strong CPU and
>maximum memory so it can also run some of the backend process such as
>spamassassin, redis or SQL database etc. (we want to start using SQL
>DB for more which means it will be under heavy use)


> What workloads can best co-exist with NFS where each does not contend for the other's resources?


NFS server uses disk and LAN,           D     L
NFS client uses RAM and LAN,              R   L
S.A. uses CPU and RAM,                    R C
Redis uses RAM and CPU                    R C
SQL uses RAM and disk and CPU           D R C 


Redis and SQL also use the lan if they're serving other servers.

> I'd put a couple smaller machines in HA in front of that to proxy
> webmail, HTTP website, IMAP and Submission. Edge MTA is on a separate
> server and would probably stay that way, maybe adding a failover. It
> keeps a fraction of its mail in the local system but will make more
> heavy use of the SQL DB which I thought to also put on the file
> server(?)



> Could NFS keep up with load for proxy of HTTP, IMAP, LDA, etc?
>
> Is local attached SATA SSD in RAID 1 ok? Will 6Gb/s SATA III be a bottleneck in any possible scenario? I was looking at motherboards with multiple PCIe or M.2 slots thinking NVMe bandwidth (3GB/s) would be great but I'm unsure if NFS, gigabit ethernet or other components could even make use of it.


see Q1

> Or is that too amateur and local attached (hope not remote attached) SAS a minimum requirement? SAS SSDs are a newer thing I never used and expensive for the $boss. I read some people express doubts that SSD is suited for SAS at all which is one reason I thought just use SATA for more efficient power and cost.
>



> Our workload:
>
> We process a lot of mail but as you can guess, don't have tremendous
> storage needs. We have several tens of thousands of users but a
> smaller fraction of that are actively using mail every day. Maildir
> storage is several TB. Exact daily mail volume is unknown but should
> be on the order of a few million, many which users have forwarded to
> to other accounts so a small fraction is stored locally.


Forwarding can get messy when gmail thinks _you_ are spamming.

SATA will probably work alright there.

> We also have few TB of web data that is hosted from a server in the
> same location that I thought to unify into the NFS setup.


I wouldn't use NFS as server disk unless the service was CPU
intensive. so servers running Redis or S.A. are good candidates for
being NFS clients. others not so much.

--
When I tried casting out nines I made a hash of it.