Howdy, Stranger!

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


Vultr New CPU's?
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.

Vultr New CPU's?

I spun up a new $10 instance today in New Jersey and noticed the virtual cpu tag was one that I haven't seen before. Anyone have any guesses on what it is?

cpu family : 6
model : 85
model name : Virtual CPU 82d9ed4018dd
stepping : 4
microcode : 0x1
cpu MHz : 2600.000
cache size : 16384 KB

gcc -c -Q -march=native --help=target | grep march
-march= knl

«1

Comments

  • FoxelVoxFoxelVox Member
    edited February 2018

    My guess is it would be an Xeon E5 2660/2670 v3 or an Xeon E3 12*0v6

  • eva2000eva2000 Veteran
    edited February 2018

    Wow interesting if that is true march = knl is Intel Knights Landing - Intel Xeon Phi based though nothing Xeon Phi based is clocked at 2.6Ghz ???

    https://en.wikipedia.org/wiki/Xeon_Phi

    what GCC version you using to determine that as it could be GCC out of date or inaccurate

    if cpu family and model are to be believed then could be Intel Xeon Gold Scalable cpus though ones with base clock of 2.6Ghz don't have 16MB cache https://ark.intel.com/compare/120487,123541,120483,123689,123548

  • The idea that it's a KNL is basically unthinkable. Those are special purpose cpus that use a lot of parallel Atom x86 cores with very fast numerics (AVX512). Where did the gcc flags come from? Can you post the cpu flags from /proc/cpuinfo? Thanks.

  • jlchandlerjlchandler Member
    edited February 2018

    @eva2000 said:
    Wow interesting if that is true march = knl is Intel Knights Landing - Intel Xeon Phi based though nothing Xeon Phi based is clocked at 2.6Ghz ???

    https://en.wikipedia.org/wiki/Xeon_Phi

    what GCC version you using to determine that as it could be GCC out of date or inaccurate

    if cpu family and model are to be believed then could be Intel Xeon Gold Scalable cpus though ones with base clock of 2.6Ghz don't have 16MB cache https://ark.intel.com/compare/120487,123541,120483,123689,123548

    Yea, it's probably old. It's the version found in the Ubuntu rep. My guess is that it's an Intel Xeon Gold 6142. Here are the flags.

    flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology eagerfpu pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm invpcid_single kaiser fsgsbase bmi1 hle avx2 smep bmi2 erms invpcid rtm avx512f clwb avx512cd xsaveopt arat

  • price per thread wise probably Intel Xeon Gold 6132 Scalable.

    Thanked by 1jlchandler
  • Here's a quick Geekbench of a $20 instance and this new cpu

    https://browser.geekbench.com/v4/cpu/6964551

    Also OpenSSL

    type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes
    md5              70235.56k   201755.40k   443285.75k   633032.01k   710647.81k
    hmac(md5)        58808.10k   175595.43k   402648.41k   571430.70k   682078.52k
    des cbc          80841.81k    84996.00k    85502.20k    85704.69k    85472.60k
    des ede3         31937.76k    32348.17k    32286.89k    32453.27k    32466.62k
    idea cbc         97242.12k   101761.48k   102762.17k   103065.77k   102719.49k
                      sign    verify    sign/s verify/s
    dsa 2048 bits 0.000287s 0.000319s   3482.0   3137.6
    

    Seems good for now. Looks a lot like Hetzner Cloud in terms of raw performance. I'm probably on an empty node tho ;)

    Thanked by 3Aidan jlchandler imok
  • FredQc said: Seems good for now. Looks a lot like Hetzner Cloud in terms of raw performance. I'm probably on an empty node tho ;)

    Looks better than Hetzner's cloud range by about ~600-1000 per core

  • Aidan said: Looks better than Hetzner's cloud range by about ~600-1000 per core

    It's about the same

    https://browser.geekbench.com/v4/cpu/compare/6629903?baseline=6964551

  • @FredQc said:

    Aidan said: Looks better than Hetzner's cloud range by about ~600-1000 per core

    It's about the same

    https://browser.geekbench.com/v4/cpu/compare/6629903?baseline=6964551

    DO is currently better.

    Re-run your Hetzner results, their CPU performance is about ~30% worse on average since their release.

    https://browser.geekbench.com/v4/cpu/6964859

    Thanked by 1willie
  • First-RootFirst-Root Member, Host Rep
    edited February 2018

    Not really surprising at that prices to be honest.

  • @Aidan said:

    @FredQc said:

    Aidan said: Looks better than Hetzner's cloud range by about ~600-1000 per core

    It's about the same

    https://browser.geekbench.com/v4/cpu/compare/6629903?baseline=6964551

    DO is currently better.

    Re-run your Hetzner results, their CPU performance is about ~30% worse on average since their release.

    https://browser.geekbench.com/v4/cpu/6964859im

    My Hetzner cloud out performed a lot of Vps companies. Pretty sure it outperformed DO by a lot unless I've confused something. I'll have to run tests later tonight at DO, Vultr.

  • XeiXei Member
    edited February 2018

    ** Edit: Vultr benchmarks are from 2-3 year old VPS I have still running. Not the new CPU mentioned here

    If I understand the results below properly, Hetzner > Vultr > DO. Both Hetzner and Vultr are drastically faster than DO. These are all using their cheapest plans, with Vultr using 2nd cheapest ($5). Note, Vultr was faster than Hetzner in the first test, I don't know what that ultimately means. I just know this was a test rm used to compare OpenSSL performance in a popular thread.

    Test:
    openssl speed -evp aes-256-cbc

    DO:
    type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes
    aes-256-cbc
    214501.92k 242644.25k 230141.43k 262390.44k 322570.27k 327221.25k

    Vultr:
    type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes
    aes-256-cbc
    541993.36k 568488.79k 587297.13k 587791.02k 586670.49k

    Hetzner:
    type 16 bytes 64 bytes 256 bytes 1024 bytes 8192 bytes 16384 bytes
    aes-256-cbc
    487307.67k 648290.58k 667913.98k 678695.90k 674627.58k 666113.37k

  • Xei said: Test: openssl speed -evp aes-256-cbc

    I wouldn't go purely by that test, which is mostly done by the AES-NI hardware. The speed of those instructions varies a lot between cpu generations. You need some wider spectrum tests. I think geekbench is not that indicative. Passmark is pretty good IME.

  • In my opinion it got better than first release.
    This isn't a benchmark (lazy fuck sorry).
    But before, on uptime I saw the cpu being used a bit even though I did practicly nothing on it. I assumed it was all my neighbores or something.

    Maybe it was all the benchers running their scripts at the same time or something.
    Now it shows 0.00 0.00 0.00

  • @willie said:

    Xei said: Test: openssl speed -evp aes-256-cbc

    I wouldn't go purely by that test, which is mostly done by the AES-NI hardware. The speed of those instructions varies a lot between cpu generations. You need some wider spectrum tests. I think geekbench is not that indicative. Passmark is pretty good IME.

    I've heard SSL is one of the most tasking for resources especially if a server has a great deal of concurrent SSL connections (high traffic server, tons of SSL handshakes, etc). Wouldn't this be the best performance test given I/O etc is great?

  • williewillie Member
    edited February 2018

    I'm in the process of compiling ffmpeg on a 2gb instance. If that takes less than 10 minutes it's doing pretty well.

    Update: 7 minutes 12 seconds. Really very very good.

  • First-RootFirst-Root Member, Host Rep
    edited February 2018

    Mind to share your exact compiler flags (codecs) so that we can compare?

  • For those of us that run high volume SSL connections is how fast something compiles more useful than the SSL test? I guess the benchmark I used is a better indicator of SSL performance? Whereas ffmpeg compile is more of an overall benchmark?

  • Not enough to do just server side testing also need to test front end metrics ie. page load times and load testing (wrk/h2load/siege/ab) over HTTP and HTTPS (http/1.1 and http/2 based). webpagetest.org and gtmetrix.com are 2 of my favs.

    Thanked by 1Xei
  • williewillie Member
    edited February 2018

    FR_Michael said: Mind to share your exact compiler flags (codecs) so that we can compare?

    I just ran ./configure from the current (as of when I ran it) snapshot and used the defaults. If you want I can spin the vm back up and run it again to get the exact flags, but I figure small differences won't change the result much.

    Xei said:

    For those of us that run high volume SSL connections is how fast something compiles more useful than the SSL test? I guess the benchmark I used is a better indicator of SSL performance?

    I think the specific SSL test you ran only tests the speed of one crypto operation, AES encryption, which is a hardware instruction on recent x86's. I think cpus are fast enough by now that cpu load from SSL isn't dominated by crypto any more. Even the slowest number you posted is 215GB/s which is 1000x the speed of any reasonable network connection you'll find on LET. I think OpenSSL even counting everything will be fairly cheap compared with your application and of talking to all those connections. But you'll have to do some careful profiling to find out.

    The raw AES speed is of more interest for things like disk encryption than SSL encryption because the bottleneck for SSL will be the application and the network.

    Added: I just ran the ffmpeg compilation on a slightly older snapshot on an i5-3470S (Kimsufi KS-4C, 3.1 ghz). 10 minutes 5 sec despite the higher cpu clock. I expect the Skylake architecture is more efficient but still this is a surprising result. I might investigate more.

    Thanked by 2Xei First-Root
  • eva2000eva2000 Veteran
    edited February 2018

    willie said: I think OpenSSL even counting everything will be fairly cheap compared with your application and of talking to all those connections. But you'll have to do some careful profiling to find out.

    Vultr bare metal with 10Gbps I managed to push to max out at 0.99GB/s too heh but yes not LET typical pricing.

    Still the differences can be measured - especially in terms of page load speed and metrics like SSL negotiation time and TTFB and speedindex/visual render time. For instance some folks upgrading to new DigitalOcean Droplet plans report worse numbers for those metrics compared to their old droplets.

    Example i benchmarked on vultr bare metal comparing

    Figuring out latency response time at idle and under load also can differentiate between VPS/server providers' offerings :)

    willie said: I expect the Skylake architecture is more efficient but still this is a surprising result. I might investigate more.

    Might want to investigate Skylake's turbo boost limits for std, avx2 and avx512 workloads as they differ so if you get some avx512 work loads, the max frequency might be alot lower https://blog.cloudflare.com/on-the-dangers-of-intels-frequency-scaling/

  • avx512 in Skylake appears to be pretty slow compared with its potential. Throughput is not really better than avx256 based on some crypto benchmarks I saw. It might improve when the 10nm cpus arrive, or might be better in KNL.

    Thanked by 1vimalware
  • JoseQuesoJoseQueso Member
    edited February 2018

    even though their cpus have gotten better lately, vultr is still terrible imo. $10 extra per month for ddos protection. when you can get buyvm's or ramnode's ddos protection for $3 lmao and countless others (virmach just off the top of my head).

    sad really. even though vultr has an insanely nice vps panel / website, fuck it, it's not worth it. tread carefully folks, and steer far far away

  • willie said: Even the slowest number you posted is 215GB/s

    MB/s not GB/s.
    You probably can't even write to L1 cache at 215GB/s.
    Also note that he's using AES-256-CBC which isn't exactly fast.

    AES-NI doesn't use >128-bit operations so won't invoke AVX slowdowns. AVX512 can still be beneficial for crypto as long as you're sticking to the AVX512VL subset.

  • xyz said: MB/s not GB

    Oops yes. Still that's for aes256 and aes128 (used in TLS) will be a fair bit faster.

  • For those interested, they're currently rolling out these new Skylake CPU's platform wide and anticipate it to be completed within the next 20 business days.

  • XeiXei Member
    edited February 2018

    Are they better than my old 3.xghz instance which I posted SSL benchmark for above? That means all new VPS are going to use the new Skylake architecture? Or you spin instances til you get one?

  • XeiXei Member
    edited February 2018

    Can someone tell me what command to use to confirm I have a Skylake CPU?

    What do I look for in /proc/cpuinfo ?

  • FredQcFredQc Member
    edited February 2018

    @Xei said:
    Can someone tell me what command to use to confirm I have a Skylake CPU?

    >

    cat /proc/cpuinfo

    What do I look for in /proc/cpuinfo ?

    Their new CPUs are rated at 2600 Mhz vs 2400 Mhz for older one

    There's also new CPU flags added like AVX512 etc

    I've tried to get the new CPU on smaller instances and it was not successful for now...

    Thanked by 1Xei
  • XeiXei Member
    edited February 2018

    Thanks so 2600Mhz is enough to confirm Skylake? What's the smallest plan to carry it so far?

Sign In or Register to comment.