Search results

You are all already using a ramdisk, https://en.wikipedia.org/wiki/Page_cache https://docs.microsoft.com/en-us/windows/desktop/fileio/file-caching

This thread is painful to read, why do people not understand that low latency will significantly increase performance. Optane is like 5000x faster than Samsung 960 Pro when running SQLite benchmarks. You know what applications that use SQLite a lot? Your web browser! And there are like billions...

There are thousands of other reasons that machine will go offline before the SSD will fail. If uptime is important, you need some kind of system redundancy. Mirroring SSDs is still a poor solution.

The Nytro SSD's looks really awesome. Another thing though, enterprise SSD's are so reliable, that mirroring them is in my opinion just a waste of performance/money and time (adds complexity). At ServeTheHome they have some statistics on Intel S3700 drives where almost noone ever failed. In...

Optane, because it will act as additional system memory. With Optane you need less expensive RAM. Especially consumer level SSD's are just slightly faster than 10k spinners. Every real benchmark where you check latency and write sync/flush they are just horrible. In these benchmarks, Optane is...

Actually it is very easy to prove this. You can use fio, dd or pg_test_fsync For example $ sudo fio --filename=/dev/nvme0n1 --direct=1 --sync=1 --rw=write --bs=4k --numjobs=10 --iodepth=1 --runtime=10 --time_based --group_reporting --name=journal-test $ sudo pg_test_fsync -f...

Still the old Intel 320 give me up 8000 iops for the same benchmarks. I would just get a SSD that gives most GB of storage per $. If you need performance, then Samsung 960 NVME does far from cut it. The truth is, most consumer level SSD's perform almost equally bad.

I would not buy Samsung 960 nvme, in my benchmarks for random write I get only 1000 iops, where with Intel 900p Optane I get 500 000 iops. That is a 500x difference!!!

This is quite common among most of the consumer grade SSD's. On Servethehome, all consumer SSD's are marked as pile of shit, crap performance, worthless and so on. On Ebay, enterprise SSD's are available for same $ per gb as consumer SSD's, but for example the Intel S3700 series still have 20x...

If you mean random read, then just get more RAM. That is where you do most of your random reads ;)

You buy enterprise SSD's from Ebay. 90% have been used for a few hours. The price per GB is the same or just slightly above consumer level SSD's

Hardware raid is only faster because it doesn't have to read/write additional io operations to disk drives for each parity block calculated. Modern CPU's with SSE3 or better can do parity calculation for thousands of MB per second. The biggest problem with hardware raid is cost though...

This is a really good answer! People should bookmark this.

Tested with pg_test_fsync and fio on Intel 6700K @ 4.6ghz on Ubuntu 17.04

They are not shit, they are okey! For spinners there is no need to upgrade.

I run raid0 on Intel S3700, they are so reliable that there is no need for mirroring, I do still have backups though!

For enterprise usage, even the good old Intel 320 is like 10 times faster than Samsung 960 Pro. For SQL Server, what matter is sync write performance, which means, write 8kb, and flush it, write another 8kb and flush it and so on, the write cache on the SSD drive doesn't really matter much. Read...

Get a 600GB Intel 320 from Ebay for $100-$120. It still has 10x better random write performance than most other consumer SSD's.

Enterprise SSDs are always WAY faster than consumer SSD's. My good old, well used Intel S3500 beats the crap out of any NVME consumer SSD. I see around 5x - 40x better performance when dealing with a lot of random io, especially writes.

Yes, try pg_test_fsync and you will se that the Samsung 960 pro fails big time.

Get enterprise grade SSD's, they are like 10-100 times faster than consumer grade SSD's for certain workloads. Consider this review with Samsung 840 Pro, which has almost identical performance to Samsung 850 Pro. It looks like a joke...

Fsync is a command to the operating system that tells it to flush data from ram to disk. This matter for a lot of things, from web browsers to virtual machines. The easiest way to benchmark this is to install Postgresql and run pg_test_fsync. You can also do it with fio or dd.

I don't use Samsung 850 EVO nor PRO, because their fsync performance is of the worst kind. My Intel S3500 has 10 timer better fsync performance.

Its not a cache but a small buffer, and all operating systems/filesystems have this. ZFS don't have any write caching. Unless you use it with hardware raid, which is expensive. I would still use ZFS though, the other filesystems may perform better with a cache, but they are also a lot more...

pgbench tests read performance. pg_test_fsync will test sync write performance

Phoronix has tested Samsung 950 and 960. SQLite performance is crap, because of horrible sync write performance. Just as I said. http://www.phoronix.com/scan.php?page=article&item=samsung-960-evo&num=2

No I'm not trolling :-) That is why I told you how to test it. You get similar results with fio. Stop using broken benchmark tools.

Nope, My old Intel 320 beats the crap out of Samsung 950 Pro or 960 for database workloads. You can easily test this with Ubuntu and PostgreSQL by running the following command: /usr/lib/postgresql/9.5/bin/pg_test_fsync Intel 320 gets between 2000-4000 IOPS. Intel S3700 should get 3000-5000...

Ram errors happen so rarely that most people shouldn't even think about it. I have plenty of machines running 24/7 that totally add over 1 TB ECC ram together, only one time during the last five years have I seen a ECC memory error. 3 times if we count the memory in a CPU.

Which may not happen a single time during our 100 years lifespan.

No filesystem will kill your filesystem if you have ram defect. There may be some metadata inconsistency, which can make recovery harder. But this is so rare, that you could as well place a bet in a lottery. Businesses do take that bet, as the cost is minor and the consequences could cost A LOT...

A ZFS scrub will NOT kill your filesystem while scrubbing if you have a memory defect. Where the hell do this come from? Who is the rat who has started spreading these rumors? Someone on the FreeNAS forum? Please those guys are idiots coming from another planet. The answer is simple, there is...

Buy enterprise stuff, like Intel S3500 or 3700. Mostly american sellers and those SSD's are much faster than most consumer SSD's. And the price is good too.

Dropbox choosed SATA drives for their new data center. By all means, SATA is good enough. While SAS is a better protocol, it is usually not worth the extra price.

What kind of random OI is this? If you have a lot, then you should consider going for a SSD build. And while we're at that, SSD's today are so reliable that you should be fine with striping, just have some backup.

Good and correct answer! BTRFS is a joke, don't even waste your time on it.

I recently benchmarked my 2TB PostgreSQL database running ZFS on Linux and FreeBSD. I got consistently 50% better performance on FreeBSD. I have also tested SQL Server on iSCSI from a FreeBSD server and 4x Intel S3500 in Raid0. I got consistently 20% better performance on read and write...

This may be because you're using consumer SSD's which an enterprise SAS controllers will perform badly on as it executes ATA_CMD_FLUSH for writes aka SAFE_WRITE.

That PNY driver probably get that write speed as it completely ignores ATA_CMD_FLUSH, which means that if something go wrong(for example a powerloss) you will be in serious pain. Most likely you will have to reinstall Windows.

No one need LSI 9300 for SATA drives. The SAS2008 chipset can do 350k iops and up to 3200MB/s