SSD VPS: Wetin E Mean and Why E Fast

SSD VPS Naa Gịnị?

SSD VPS na abụ virtual private server wey disk ya na-eji flash memory, kama spinning hard disk. SSD pụtara solid state drive: storage wey NAND flash chips rụnyere, enweghị moving parts n'ime ya. VPS pụtara virtual private server: otu isolated slice nke physical host machine, wey na-agba operating system nke ya, a na-erekwa ya nye gị ka ọ dị ka whole server. Tinye ha abụọ ọnụ, label ahụ na-ekwe nkwa otu ihe. Mgbe server gị na-agụ data block, ọ dịghị mechanical part ga-ebu ụzọ kwaga.

Nke ahụ bụ nkọwa ya niile. Ihe niile dị n'okpuru na-akọ maka ihe label ahụ anaghị ekwu, maka na okwu "SSD hosting" nọ na pricing page anaghị agwa gị ọtụtụ ihe gbasara storage array wey dị n'azụ ya.

Ọ bụrụ na ị ka na-achọpụta otú VPS si dị iche na plain virtual machine, ọdịiche dị n'etiti VPS, VM na VPC bụ page ka mma ịgụ na mbụ.

Why hosts dey advertise storage speed instead of storage size

A plan page dey list CPU cores, memory, disk size and bandwidth, then e add one word about the disk wey no be size at all. Hosts dey do am because capacity stop to be the important storage number many years ago. The two numbers wey decide how server go feel na IOPS (input/output operations per second) and latency (how long one operation dey take to come back).

The mechanical difference na the reason. Hard disk dey keep data for rotating platters and dey read am with head wey dey on moving arm. To reach block for another place on the platter, the arm gatz seek, then the disk gatz turn until the correct sector reach under the head. For 7200 revolutions per minute, half turn dey take about 4 ms on average, and the seek dey add several more. Flash no get arm or platter, so read na electrical lookup wey dey return within tens of microseconds.

[
  {
    "device": "7200 rpm hard disk",
    "random_read_iops": "125",
    "read_latency_ms": 8
  },
  {
    "device": "SATA SSD",
    "random_read_iops": "90,000",
    "read_latency_ms": 0.15
  },
  {
    "device": "NVMe SSD",
    "random_read_iops": "600,000",
    "read_latency_ms": 0.08
  }
]

A 7200 rpm hard disk get rating of roughly 125 random 4k reads per second, and one of those reads dey take about 8 ms to come back. An NVMe drive get rating near 600,000 of the same reads, at about 0.08 ms each. The SATA SSD wey dey between dem get rating near 90,000. Read those figures as orders of magnitude, no be percentage improvement.

Two warnings dey about that data. These na published specifications for complete drives for each class, so dem be manufacturer sheet figures, and no be something wey dem measure for VPS. And you no dey ever get complete drive: your volume na share of one device, or share of one array, wey dey beside other customers for the same hardware.

The two columns dey answer different questions, so read dem together. Latency na how long you dey wait for one operation. IOPS na how many operations the drive fit work on at the same time. Flash dey reach high IOPS figures through parallelism, because many flash chips dey answer many requests at once while deep queue dey keep dem busy. A single threaded program wey dey issue one read, wait for am, then issue the next one, no go ever see the top figure for that chart. E go see the latency column instead.

SSD, NVMe, SATA and PCIe: okwu anọ maka layers anọ dị iche

Buyers dey confuse okwu ndị a n’ihi na nke ọ bụla na-akọwa akụkụ dị iche nke system.

• SSD na-akọwa medium. E mean say data dey NAND flash chips, kama ọ dey magnetic platters.
• SATA na-akọwa interface, dem design am for the time wey mechanical disks dey common. E dey reach only 6 Gbit/s, wey be around 550 MB/s real throughput, and e command queue fit hold 32 outstanding commands.
• NVMe (non-volatile memory express) na protocol wey dem write specially for flash. E support many queues, and queue ọ bụla fit hold thousands of commands. So several CPU cores fit talk to drive at the same time without sharing one narrow queue.
• PCIe (peripheral component interconnect express) na bus wey NVMe dey run on. Na the same kind lanes wey graphics card dey plug into.

So SATA SSD and NVMe SSD both dey store data for flash. The difference na the interface wey each one dey use talk. SATA SSD still dey far quicker than any hard disk, but that 32 command queue dey limit how much parallel work e fit express. Parallel work na exactly wetin flash good for. Which one worth paying for depend on your workload, and the decision between NVMe and SATA SSD explain the matter properly.

Local flash or network attached storage?

Two different arrangements dey sell under the same word.

Local storage mean say the flash drives dey inside the same physical host wey your VPS dey use. Request dey travel through PCIe inside one machine and come back directly. Na so latency fit stay for tens of microseconds.

Network attached storage mean say your virtual disk dey for another storage cluster, often Ceph or a SAN (storage area network). Every read and write dey cross network before e reach the disk. Providers usually dey call am "cloud block storage" or "elastic volumes". The flash dey real. The network trip dey real too, and dem add am to every operation. So latency fit reach high hundreds of microseconds or low milliseconds instead of tens of microseconds.

None of them be the wrong answer. Network storage fit survive when host fail, because the data no dey that host before: provider fit start your server for another hardware and the disk go follow am. Local NVMe dey faster and e dey tied to one physical machine. So if hardware fail there, you need restore from backup. Ask which one the plan dey use. Almost nobody dey ask.

How di difference dey feel for real server

Server work mostly na small random reads and writes, no be long sequential transfers. Na why the MB/s figure for marketing copy dey page na the number wey useful pass.

• Database commits. Database wey promise durability dey call fsync when transaction commit, then e dey wait for drive to confirm say data don really store. For hard disk, that wait na milliseconds, and e dey hold small database to low hundreds of commits per second. For flash, the same wait na fraction of one millisecond. Na here the gap wide pass, whether you dey run PostgreSQL, MySQL or SQLite as production database.
• Package installs. apt install dey unpack thousands of small files and sync dem to disk as e dey go. Almost none of this work dey sequential, so IOPS na wetin dey limit am.
• Container image pulls. docker pull dey fetch compressed layers through network, then e dey extract dem into thousands of small files. Network dey limit the download. Disk dey limit the extraction, and for slow volume, na the extraction dey make you sit down dey wait.
• Boot and reboot. Startup dey read kernel and initramfs, then hundreds of small unit files and shared libraries wey scatter across the volume.

None of these na large sequential read. Volume wey dey stream 500 MB/s but dey serve only 3,000 IOPS go still feel slow during docker compose pull, because dem dey count the wait for each file, no be for each megabyte.

Why "SSD cloud hosting" for pricing page no dey tell you almost anything

The word only describe the medium and stop there. E no tell you anything about the interface wey the drive dey use, and e no tell you whether the drive dey inside the same machine as your server. E still no tell you the maximum level wey your plan fit reach.

That maximum level matter pass, but na the one providers advertise least. Providers dey limit IOPS and throughput for each volume, because one host dey serve many customers and one neighbour without limit fit use all the resources, leaving the others without enough. Limit of some thousand IOPS for hardware wey fit handle hundreds of thousands normal and honest, but e no dey visible for the plan description. Two plans fit both say "SSD", while one na local NVMe without per volume limit and the other na shared cluster volume limited to 3,000 IOPS.

Limits get two forms. Sustained limit na fixed maximum wey no dey change. Burst limit give you low baseline plus credits wey let you pass the limit for some time, and the credits dey refill when the volume idle. Burst limit fit look excellent for five minute test, then e go drop to the baseline during database import or big restore. If provider quote big number, ask how long you fit maintain am.

How to check wetin your VPS really give you

From inside the guest, you fit only see wetin the hypervisor tell you.

lsblk -d -o NAME,ROTA,MODEL,SIZE
cat /sys/block/vda/queue/rotational

Replace vda with the device name lsblk wey e print for your disk. The ROTA and rotational file read 0 when dem don tell the kernel say the device no dey rotate, and 1 when dem don tell am the opposite. Virtual disk dey set that flag from wetin the hypervisor advertise, so e describe the virtual device, no be the physical hardware underneath am. MODEL dey usually empty for virtio disk like vda, or e fit be generic string like QEMU HARDDISK for emulated SATA controller. Guests no suppose see the host array, and dem no dey see am.

So treat the flag as hint and measure the remaining things. Run random 4k test with fio against file for the volume, with the queue depth wey your application really dey use, and run am long enough to use any burst credit. How to benchmark VPS properly get the fio commands and the common mistakes wey dey produce numbers wey look better than reality.

Dree questions to ask provider before you buy

1. Storage dey local to the hypervisor, or e dey network attached? The answer go set the lowest latency wey you fit expect, and e go decide wetin go happen to your data if physical host fail. Provider wey answer this one plainly don think about am.
2. Wetin be the IOPS limit for my volume? Ask for number. "Unlimited" and "enterprise grade" no be numbers. If truly no cap dey, ask wetin dey stop neighbour for the same host from using the whole array during their backup window.
3. That limit na sustained or burst? If na burst, ask for the baseline and how long the burst fit last. The baseline na the figure wey your nightly job go work with, so na am you need plan around.

Flash go wear out, and na your problem?

Flash cells fit only limited number of write cycles, so drives dey publish endurance rating for TBW (terabytes written) or DWPD (drive writes per day). The drive dey spread writes evenly across its cells; dem dey call this wear levelling. E also keep spare blocks for reserve to replace cells wey fail. For VPS, the wear na provider problem: dem dey monitor SMART counters for their own drives and replace hardware before the rating finish. Your problem older pass flash. Drive no be backup, and redundancy under your volume no be backup too, because e dey copy deletion just as faithfully as e dey copy data.

The real trade you dey make na price per gigabyte. Flash cost more per gigabyte than spinning disks, so SSD plan usually get less capacity than hard disk plan for the same price. If you need plenty space for media or archives, keep the fast volume small and put the bulk for somewhere cheaper. Na the correct arrangement for keeping backups off the server too. To see how storage fit inside the rest of the bill, wetin VPS really cost dey break down the line items.

FAQ

SSD VPS dey be the same thing as NVMe VPS?

Every NVMe VPS na SSD VPS, because NVMe drives na flash. But the reverse no hold. Plan wey dem advertise as "SSD" fit be SATA SSD. Na flash behind interface wey dem design for mechanical disks, with 32 command queue and ceiling near 550 MB/s. Both dey far faster than hard disk. If the difference between dem matter for your workload, ask the provider which one the plan dey use instead of judging am from the plan name.

SSD VPS go make my website faster?

E go make disk work faster, and na only disk work. Page wey dey run several database queries for each request go faster, because those queries and their commits na small random I/O. Page wey memory or cache dey serve never touch disk as e dey go out, so e go change very little. Measure which part of the request slow before you pay for storage to fix am.

How I fit check whether my VPS really dey use SSD storage?

From inside the guest, you no fit verify the physical hardware. lsblk -d -o NAME,ROTA shows wetin the virtual device advertise, and the hypervisor dey decide that value, so 0 there na hint, no be proof. The practical check na measurement: run fio with random 4k read workload for several minutes, then check the latency wey e report. Single digit millisecond random reads point to spinning disks or congested network volume. Tens of microseconds point to local flash.

Network attached SSD storage worse pass local NVMe?

E slower for each operation, and e dey fail differently. Network round trip dey add to every read and every write, so latency dey higher even though both na flash. In exchange, your data no dey stay on one physical host, so host failure no go carry the volume along, and snapshots plus live migration easier for the provider. Choose local flash for database wey latency dey important. Choose network storage when survival of the volume matter pass microseconds.