PCIe SSD, SATA SSD or Hard disk drives for a 10-workers server?

Question

For a server aimed to centralize the documents of a 10-people team, which kind of media would you advise to have a system that is both responsive and reliable? I want the server not hanging and that when opening documents users have the same impression as if those were stored on their own desktop.

I see the following benefits and drawbacks:

A) PCIe SSD

• fastest transfer rates
• very pricey

B) SATA SSD

• fast (but not as much as PCIe however)
• existing models for datacenters
• could be used in RAID 1 for full redundancy
• limited capacity

C) HDD with large buffer

• low cost
• available in large capacities
• could be used in RAID 1 (or in RAID 1+0 to increase transfer speeds)
• slower than SSD

Having two datacenter-grade SSD (Intel DC S3500 480 GB), I was planning using either only one of them, or both in RAID 1. Would it be a good choice for a small-business server?

Assuming that archives would be stored on larger hard drives, I plan using the SSD for the OS.

Is there a way to capitalize on the free remaining space of SSD as some kind of buffer, in the same spirit as how SSHD (hybrid drives) work?

Answer 1

First of all, impression as if stored on their own desktop is not something you will be able to get, unless files are very small. Accessing files via network is, of course, noticeably slower. Not much, but just noticeably. Though iSCSI sometimes does an awesome job at caching locally (SMB not so).

Second, a simple SATA hard disk (with nothing special) is almost certainly entirely sufficient. Not only will the server perform caching in RAM, but also the harddisk is not the bottleneck. 10 workers is not much of a challenge to even a small NAS intended for home use, unless what they call "work" is something like transcoding video streams. 10 people working, that means maybe two or three dozen disk accesses per minute, at most. What about it?

On a GbE network, you can get around 110-115 MB/s throughput, not more (well, twice that, if your switch as well as your computers support link aggregation). A single harddisk can deliver that. Easily. While this sounds good, remember that you will not always have large contiguous reads. In fact, most of the time you don't.
Access time over network is, well, let's say not that great. So, whether or not the actual disk is somewhat slower or faster makes not that much of a difference.

We're using both SMB shares and iSCSI on Synology over GbE here (with link aggregation support), and the disk is by no means the limiting factor. I've tried every setup from a single disk to 4-disk RAID-0 (as well as every other RAID level that makes sense), and the real, observable difference is exactly zero.

I haven't tried SSD, but I am very confident that the difference would be zero, too.

To give you an idea about how much latency can matter (not raw troughput), if you compare a task that is prone to access time such as for example indexing a 12GB movie in Avidemux (lots of small reads), I get about 300 frames per second indexed over SMB, about 650 over iSCSI, about 20,000 and 25,000-27,000 on SATA-SSD and nvme-SSD, respectively (both Samsung 950s). With the movie being in the buffer cache, I get 33,000-34,000. That's network versus local, and bus versus no bus.