Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!


Contabo VPS S NVMe: isn't this too slow?
New on LowEndTalk? Please Register and read our Community Rules.

All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.

Contabo VPS S NVMe: isn't this too slow?

I have a Contabo VPS S with the NVMe addon and I am benchmarking writing to the disk.

This is the speed I get:

dd if=/dev/urandom of=/dev/sda bs=4M conf=fsync status=progress
18975031296 bytes (19 GB, 18 GiB) copied, 469 s, 40.4 MB/s 

It does not go higher than 40MB/s.
Isn't this too slow for the advertised NVMe?

Comments

  • ehabehab Member

    for NVMe or ssd do a

    ./yabs.sh -i -g -r

    and show us your penis numbers .

  • @ehab said:
    for NVMe or ssd do a

    ./yabs.sh -i -g -r

    and show us your penis numbers .

    i got a 7.5

    Thanked by 2ehab cybertech
  • ehabehab Member

    @SirFoxy said:
    i got a 7.5

    i didn't see that coming

    hahahhahahahha

  • Probably CPU bottleneck, try a faster way of generating random data to test with

  • ehabehab Member

    @darkimmortal said:
    Probably CPU bottleneck, try a faster way of generating random data to test with

    do you like it fast.

  • adnsadns Member

    The 40 MB/s rw speed is usualy in 4K IOPS. That written uper, it seems to be a CPU limitation. If I understand you well, it is a new VPS. In my experience the newer VPSs has very poor performance. I bought a VPS M NVMe in BF. When I bought steal is 0.0, 512K IOPS and 1M IOPS near 20 GB/s in total. GB single core circa 800 and multi 4800. Next day the speeds average 5 GB/s, and CPU cores 600/2200. I reported the issue (poor performance), the ISP solved it for a short period, and it started again. I cancelled the service. The normal VPS packages has a good trade-of witht the friendly prices but the NVMe is a more expensive furthermore in practice unusable by the choppy speeds and CPU performance.

  • It is actually not a new one but I reinstalled it.
    I thought /dev/urandom would not use much CPU...but here is YABS output:

    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    #              Yet-Another-Bench-Script              #
    #                     v2022-02-18                    #
    # https://github.com/masonr/yet-another-bench-script #
    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    
    Sat 12 Mar 2022 09:36:16 PM CET
    
    Basic System Information:
    ---------------------------------
    Processor  : AMD EPYC 7282 16-Core Processor
    CPU cores  : 4 @ 2794.748 MHz
    AES-NI     : ✔ Enabled
    VM-x/AMD-V : ❌ Disabled
    RAM        : 7.8 GiB
    Swap       : 0.0 KiB
    Disk       : 97.9 GiB
    
    Preparing system for disk tests..../yabs.sh: line 450: curl: command not found
    Fio binary download failed. Running dd test as fallback....
    
    dd Sequential Disk Speed Tests:
    ---------------------------------
           | Test 1      | Test 2      | Test 3      | Avg        
           |             |             |             |            
    Write  | 206 MB/s    | 318 MB/s    | 293 MB/s    | 272.33 MB/s
    Read   | 482 MB/s    | 593 MB/s    | 680 MB/s    | 585.00 MB/s
    
    
    Thanked by 1ehab
  • ehabehab Member
    edited March 2022

    @deer76 said:

    can you please install

      curl
      fio
    

    and run again the yabs

  • jarjar Patron Provider, Top Host, Veteran

    It's only too slow if it bottlenecks the actual use case. My raspberry pi setups fall every benchmarking standard for LET but they sure do wonderfully handle their jobs.

  • 99% of users 40mb/s are fully enough. It is like a car - how often you drive your car at a full speed?

  • ehabehab Member

    @Btclists said:
    99% of users 40mb/s are fully enough. It is like a car - how often you drive your car at a full speed?

    i like it fast

  • @ehab said:
    can you please install

      curl
      fio
    

    and run again the yabs

    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    #              Yet-Another-Bench-Script              #
    #                     v2022-02-18                    #
    # https://github.com/masonr/yet-another-bench-script #
    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    
    Sun 13 Mar 2022 12:07:49 AM CET
    
    Basic System Information:
    ---------------------------------
    Processor  : AMD EPYC 7282 16-Core Processor
    CPU cores  : 4 @ 2794.748 MHz
    AES-NI     : ✔ Enabled
    VM-x/AMD-V : ❌ Disabled
    RAM        : 7.8 GiB
    Swap       : 0.0 KiB
    Disk       : 97.9 GiB
    
    fio Disk Speed Tests (Mixed R/W 50/50):
    ---------------------------------
    Block Size | 4k            (IOPS) | 64k           (IOPS)
      ------   | ---            ----  | ----           ---- 
    Read       | 23.31 MB/s    (5.8k) | 368.37 MB/s   (5.7k)
    Write      | 23.32 MB/s    (5.8k) | 370.31 MB/s   (5.7k)
    Total      | 46.64 MB/s   (11.6k) | 738.69 MB/s  (11.5k)
               |                      |                     
    Block Size | 512k          (IOPS) | 1m            (IOPS)
      ------   | ---            ----  | ----           ---- 
    Read       | 5.12 GB/s    (10.0k) | 3.08 GB/s     (3.0k)
    Write      | 5.39 GB/s    (10.5k) | 3.28 GB/s     (3.2k)
    Total      | 10.51 GB/s   (20.5k) | 6.37 GB/s     (6.2k)
    
  • @deer76 said:

    @ehab said:
    can you please install

      curl
      fio
    

    and run again the yabs

    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    #              Yet-Another-Bench-Script              #
    #                     v2022-02-18                    #
    # https://github.com/masonr/yet-another-bench-script #
    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    
    Sun 13 Mar 2022 12:07:49 AM CET
    
    Basic System Information:
    ---------------------------------
    Processor  : AMD EPYC 7282 16-Core Processor
    CPU cores  : 4 @ 2794.748 MHz
    AES-NI     : ✔ Enabled
    VM-x/AMD-V : ❌ Disabled
    RAM        : 7.8 GiB
    Swap       : 0.0 KiB
    Disk       : 97.9 GiB
    
    fio Disk Speed Tests (Mixed R/W 50/50):
    ---------------------------------
    Block Size | 4k            (IOPS) | 64k           (IOPS)
      ------   | ---            ----  | ----           ---- 
    Read       | 23.31 MB/s    (5.8k) | 368.37 MB/s   (5.7k)
    Write      | 23.32 MB/s    (5.8k) | 370.31 MB/s   (5.7k)
    Total      | 46.64 MB/s   (11.6k) | 738.69 MB/s  (11.5k)
               |                      |                     
    Block Size | 512k          (IOPS) | 1m            (IOPS)
      ------   | ---            ----  | ----           ---- 
    Read       | 5.12 GB/s    (10.0k) | 3.08 GB/s     (3.0k)
    Write      | 5.39 GB/s    (10.5k) | 3.28 GB/s     (3.2k)
    Total      | 10.51 GB/s   (20.5k) | 6.37 GB/s     (6.2k)
    

    this looks really good actually

    Thanked by 1lentro
  • jarjar Patron Provider, Top Host, Veteran

    @ehab said:

    @Btclists said:
    99% of users 40mb/s are fully enough. It is like a car - how often you drive your car at a full speed?

    i like it fast

    I got a fast, hard disk for you.

  • jsgjsg Member, Resident Benchmarker

    @deer76 said:
    18975031296 bytes (19 GB, 18 GiB) copied, 469 s, 40.4 MB/s

    That's actually not bad for fsync'd writing. Also keep in mind that writing out 19 GB is likely to trash caching and in particular the NVMe on-board cache.

    @darkimmortal said:
    Probably CPU bottleneck, try a faster way of generating random data to test with

    No. While creating fat pseudo random streams does indeed need some CPU performance (a) that's still orders of magnitude faster than any flash storage, and (b) we are talking about a system with 4 modern Zen vCores.

    That said, Contabo's forte is not speed, neither CPU (Ryzens are way faster), nor disk (even a few abusers make it considerably slower for all), nor network (it's quite OK but not really fast). Contabo's forte is selling quite decent quite large VPS on modern hardware for a low price (for what you get).

    Thanked by 1deer76
  • Alright, thanks to everyone that was on topic!

  • ehabehab Member

    @jar said:
    I got a fast, hard disk for you.

    i know you always deliver

    Thanked by 1jar
  • darkimmortaldarkimmortal Member
    edited March 2022

    @deer76 said:
    It is actually not a new one but I reinstalled it.
    I thought /dev/urandom would not use much CPU...but here is YABS output:

    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    #              Yet-Another-Bench-Script              #
    #                     v2022-02-18                    #
    # https://github.com/masonr/yet-another-bench-script #
    # ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## ## #
    
    Sat 12 Mar 2022 09:36:16 PM CET
    
    Basic System Information:
    ---------------------------------
    Processor  : AMD EPYC 7282 16-Core Processor
    CPU cores  : 4 @ 2794.748 MHz
    AES-NI     : ✔ Enabled
    VM-x/AMD-V : ❌ Disabled
    RAM        : 7.8 GiB
    Swap       : 0.0 KiB
    Disk       : 97.9 GiB
    
    Preparing system for disk tests..../yabs.sh: line 450: curl: command not found
    Fio binary download failed. Running dd test as fallback....
    
    dd Sequential Disk Speed Tests:
    ---------------------------------
           | Test 1      | Test 2      | Test 3      | Avg        
           |             |             |             |            
    Write  | 206 MB/s    | 318 MB/s    | 293 MB/s    | 272.33 MB/s
    Read   | 482 MB/s    | 593 MB/s    | 680 MB/s    | 585.00 MB/s
    
    

    @jsg said:

    @deer76 said:
    18975031296 bytes (19 GB, 18 GiB) copied, 469 s, 40.4 MB/s

    That's actually not bad for fsync'd writing. Also keep in mind that writing out 19 GB is likely to trash caching and in particular the NVMe on-board cache.

    @darkimmortal said:
    Probably CPU bottleneck, try a faster way of generating random data to test with

    No. While creating fat pseudo random streams does indeed need some CPU performance (a) that's still orders of magnitude faster than any flash storage, and (b) we are talking about a system with 4 modern Zen vCores.

    That said, Contabo's forte is not speed, neither CPU (Ryzens are way faster), nor disk (even a few abusers make it considerably slower for all), nor network (it's quite OK but not really fast). Contabo's forte is selling quite decent quite large VPS on modern hardware for a low price (for what you get).

    /dev/urandom really is that slow on modern CPUs, try it yourself.

    E.g. getting numbers like this on Zen and Skylake:

     dd if=/dev/urandom of=/dev/null bs=4M status=progress                
    419430400 bytes (419 MB, 400 MiB) copied, 9 s, 46.4 MB/s^C
    108+0 records in
    107+0 records out
    448790528 bytes (449 MB, 428 MiB) copied, 9.76465 s, 46.0 MB/s
    

    The conv=fsync is mostly irrelevant, it only syncs once on completion, not between each record

  • jsgjsg Member, Resident Benchmarker

    @darkimmortal said:

    @darkimmortal said:
    Probably CPU bottleneck, try a faster way of generating random data to test with

    No. While creating fat pseudo random streams does indeed need some CPU performance (a) that's still orders of magnitude faster than any flash storage, and (b) we are talking about a system with 4 modern Zen vCores.

    /dev/urandom really is that slow on modern CPUs, try it yourself.

    E.g. getting numbers like this on Zen and Skylake:

     dd if=/dev/urandom of=/dev/null bs=4M status=progress                
    419430400 bytes (419 MB, 400 MiB) copied, 9 s, 46.4 MB/s^C
    108+0 records in
    107+0 records out
    448790528 bytes (449 MB, 428 MiB) copied, 9.76465 s, 46.0 MB/s
    

    The conv=fsync is mostly irrelevant, it only syncs once on completion, not between each record

    I did and you are partly right. On linux the command dd if=/dev/urandom of=/dev/null bs=4M count=256 status=progressis indeed crap-slow, on a current FreeBSD though I get about 5 times higher performance!

    Also it seems that the whole thing is kind of weird because I get significantly better (ca. +25%) performance on an E5 v2 than on a Ryzen 39xxX.

    Regarding fsync you are right though. dd indeed only seems to do a sync once at the end rather than after each block as I assumed. That said, writing to /dev/zero fsync is kind of moot anyway.

    FWIW I also ran exactly the same dd command on my Contabo NVMe VPS and got 146 MB/s (on FreeBSD though) so @deer76's result seems plausible and in fact even good because on linux my result probably would be lower than his.

  • I should have used oflag=sync instead.

  • What is the normal speed of SSD for contabo? After I applied to unlock the SSD speed, I got the following data, which doesn't feel very fast

    fio Disk Speed Tests (Mixed R/W 50/50):

    Block Size 4k (IOPS) 64k (IOPS)
    Read 5.29 MB/s (1.3k) 64.76 MB/s (1.0k)
    Write 5.32 MB/s (1.3k) 65.15 MB/s (1.0k)
    Total 10.61 MB/s (2.6k) 129.92 MB/s (2.0k)
    Block Size 512k (IOPS) 1m (IOPS)
    ------ --- ---- ---- ----
    Read 127.62 MB/s (249) 97.88 MB/s (95)
    Write 134.40 MB/s (262) 104.40 MB/s (101)
    Total 262.02 MB/s (511) 202.29 MB/s (196)
  • @SuperValueHost said:
    What is the normal speed of SSD for contabo? After I applied to unlock the SSD speed, I got the following data, which doesn't feel very fast

    fio Disk Speed Tests (Mixed R/W 50/50):

    Block Size 4k (IOPS) 64k (IOPS)
    Read 5.29 MB/s (1.3k) 64.76 MB/s (1.0k)
    Write 5.32 MB/s (1.3k) 65.15 MB/s (1.0k)
    Total 10.61 MB/s (2.6k) 129.92 MB/s (2.0k)
    Block Size 512k (IOPS) 1m (IOPS)
    ------ --- ---- ---- ----
    Read 127.62 MB/s (249) 97.88 MB/s (95)
    Write 134.40 MB/s (262) 104.40 MB/s (101)
    Total 262.02 MB/s (511) 202.29 MB/s (196)

    That's still too slow.

    The speed should be similar to this one after the I/O limit is lifted.

    fio Disk Speed Tests (Mixed R/W 50/50):
    ---------------------------------
    Block Size | 4k            (IOPS) | 64k           (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 5.30 MB/s     (1.3k) | 64.10 MB/s    (1.0k)
    Write      | 5.32 MB/s     (1.3k) | 64.52 MB/s    (1.0k)
    Total      | 10.62 MB/s    (2.6k) | 128.62 MB/s   (2.0k)
               |                      |
    Block Size | 512k          (IOPS) | 1m            (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 524.08 MB/s   (1.0k) | 632.26 MB/s    (617)
    Write      | 551.93 MB/s   (1.0k) | 674.37 MB/s    (658)
    Total      | 1.07 GB/s     (2.1k) | 1.30 GB/s     (1.2k)
    
  • SpritiSpriti Member
    edited March 2022

    @SuperValueHost @Leftup
    I can confirm the values of SuperValueHosts. Test it on 6 machines, each on an other host, some in Düsseldorf, some in Nürnberg, different sizes (S, L, M, XL)

    Only 1 Host has higher disk performance (4x times more at 4K, 2x at 64k, 512k & 1m the same values). That host, i remember, we had performance problems at the beginning and after Mail Contabo they do something (set higher limits??? don't know).

    Be honest, Contabos SSD performance is not the best, for that you have RAM to cache files. You get what you pay for, but price / performance is still great.

    I hope NVMe has the performance that other providers SSD have, soon i will order a further server with NVMe and can test it.

    Here is a comparission with a Netcup RS 8000 SSD G7 SE 12M (~ 6 Years old) E5-2680 V4 @2.4 GHz

    fio Disk Speed Tests (Mixed R/W 50/50):
    ---------------------------------
    Block Size | 4k            (IOPS) | 64k           (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 60.68 MB/s   (15.1k) | 619.46 MB/s   (9.6k)
    Write      | 60.52 MB/s   (15.1k) | 622.19 MB/s   (9.7k)
    Total      | 121.20 MB/s  (30.3k) | 1.24 GB/s    (19.4k)
               |                      |
    Block Size | 512k          (IOPS) | 1m            (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 1.09 GB/s     (2.1k) | 1.28 GB/s     (1.2k)
    Write      | 1.04 GB/s     (2.0k) | 1.20 GB/s     (1.1k)
    Total      | 2.13 GB/s     (4.1k) | 2.48 GB/s     (2.4k)
    
  • @Spriti said:
    @SuperValueHost @Leftup
    I can confirm the values of SuperValueHosts. Test it on 6 machines, each on an other host, some in Düsseldorf, some in Nürnberg, different sizes (S, L, M, XL)

    Only 1 Host has higher disk performance (4x times more at 4K, 2x at 64k, 512k & 1m the same values). That host, i remember, we had performance problems at the beginning and after Mail Contabo they do something (set higher limits??? don't know).

    Be honest, Contabos SSD performance is not the best, for that you have RAM to cache files. You get what you pay for, but price / performance is still great.

    I hope NVMe has the performance that other providers SSD have, soon i will order a further server with NVMe and can test it.

    Here is a comparission with a Netcup RS 8000 SSD G7 SE 12M (~ 6 Years old) E5-2680 V4 @2.4 GHz

    fio Disk Speed Tests (Mixed R/W 50/50):
    ---------------------------------
    Block Size | 4k            (IOPS) | 64k           (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 60.68 MB/s   (15.1k) | 619.46 MB/s   (9.6k)
    Write      | 60.52 MB/s   (15.1k) | 622.19 MB/s   (9.7k)
    Total      | 121.20 MB/s  (30.3k) | 1.24 GB/s    (19.4k)
               |                      |
    Block Size | 512k          (IOPS) | 1m            (IOPS)
      ------   | ---            ----  | ----           ----
    Read       | 1.09 GB/s     (2.1k) | 1.28 GB/s     (1.2k)
    Write      | 1.04 GB/s     (2.0k) | 1.20 GB/s     (1.1k)
    Total      | 2.13 GB/s     (4.1k) | 2.48 GB/s     (2.4k)
    

    Indeed you need to open ticket (via email) to adjust the I/O limit. But some people says that the limit come back after some times. I never experienced it though.

    For me 4 vCPU, 8GB RAM and 400GB SSD in Singapore for $10 USD is still a good deal.
    I used it for development machine, so the performance requirements not that high.

    Thanked by 1Spriti
  • Indeed it's a little slow.

  • In fact, I have open a ticket to adjust the IO speed limit, but I still feel very slow, so I just ask if such ssd speed is normal at contabo

  • I have used the singapore server from India. It can sure take a lot of loads but the speed is not great. Its under normal. The CPU must be shared too much as they are having a seperate dedicated virtual machines, at a 3x price!!.

    At the end of the day we get what we paid for..thats it. I do not host a website that relies on a speed, Instead I used it for my inventory management that need a high loads but doesnt necessarily be fast.

  • @Btclists said:
    99% of users 40mb/s are fully enough. It is like a car - how often you drive your car at a full speed?

    Yes, I have dedis with 30MB/s I/O speed, it's more than usable, though, this is SCSI.

Sign In or Register to comment.