Page 6 of 10 FirstFirst ... 3456789 ... LastLast
Results 126 to 150 of 243
  1. #126
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    As I stated above, we've been monitoring for over 8 weeks - with no degradation.

    How, when the HV is 'full' and the SSD is 'full' would performance degrade over time?
      0 Not allowed!

  2. #127
    Rewriting data on SSD will have like 10-20% of initial performance when the drive is nearly-full because it's very costly to reinitialize the cell which otherwise is handled by TRIM.

    Under normal circumstances, re-writing the data is just picking a free cell + write and the old cell gets re-initialized in the background. When there are no erased cells drive needs to erase some before it can write.

    You can see this behaviour for example, on here:
    http://www.anandtech.com/show/6161/i...rds-we-test-it
      0 Not allowed!

  3. #128
    Join Date
    Nov 2000
    Location
    localhost
    Posts
    3,771
    Unless you a minimum SLA like EC2 provision IOPS does - excluding a defined and minimal tolerance - it is just as classy (or worthwhile, shame about the money you spent though) as the solus-in-a-box hosts that rent a dedicated server with SSD and then get their equally clueless friends or first customers to do a `dd` benchmark and then run to WHT/LEB to post their whack results.

    Over time more guests will join the node, in addition those existing users will move from testing/developing to real load scenarios and most probably as their respectives sites/webapps/services grow they will become more resource intensive and thus degrade the performance for other guests in the absence of any well tested policies.

    Useless you have policies to enforce a minimum SLA in face other demanding guests then you havent got a hope in hell, the fact you mention monitoring for 8 weeks suggest you are in this boat. Why didnt you fill the node with 35 artifical guests in the lab and simulate potential IO demands (plus some more), and see how predictable and consistent performance is and if you are able to keep it within a define range for SLA purposes

    Sounds to me like you're throwing a lot of money at it, best tech etc.. but still chancing it, no SLA with regards to bandwidth/iops etc..."monitoring" looking good etc..

    slawek makes good points too..
    Last edited by MattF; 11-13-2012 at 08:02 AM.
    MattF - Since the start..
      0 Not allowed!

  4. #129
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    When I said 'full' I mean from a contention / customer density point of view, not capacity.

    MattF - my point is that we've redesigned to achieve these numbers -at- capacity (clearly I'm not explaining this clearly! )

    We had the architecture in the lab with simulated loads for over 6 months, which is why it's now live
      0 Not allowed!

  5. #130
    When I said 'full' I mean from a contention / customer density point of view, not capacity.
    Ah, i thought you asked why SSD throughput can degrade (which could be of course 100% prevented by user).

    As for this question, just because you didn't do math, you only have some synthetic benchmarks running in the lab, yes, that always look nice as Matt said But the other user pointed out that when you add actual numbers the worst case scenario is actually 60MBps if you won't decide to oversell later.

    My point is that maybe instead of having lab full of hardware, maybe sometimes it's better to use a calculator

    @MattF: Amazon is just a pile of BS. Guaranteed IOPS are worth nothing when nodes itself are slow and unstable (you can get a preety fast, new CPU or 5 years old **** as same node, randomly). Besides that IOPS quarantee is probably no better than their whole ******, overpriced service.

    I can at least recommend dediserve for small sites, from my personal opinion amazon is good for nothing. I even wonder why so many people are using it - for me it was horrible. Nodes are overpriced to HELL, and the free offer is so slow that it's UNUSABLE. I'd be more happy with mid-range server for $200 than with $4k worth of unstable services bought at Amazon

    @Dedi, im not saying your service is bad. You just advertise for WRONG part of the market. Shared cloud can only work for small sites proper path should be:
    shared hosting -> vps -> shared cloud -> private cloud or bare metal.

    Of course if we won't took you BS scenarios into the equation (that some day somehow i got featured in TV and i just switch the sliders from leftmost to rightmost, yeah, very likely to happen ). Well too bad, in recent 5 years i didn't get "featured" so much that my servers would die That probably saved me like $200-300k USD, probably even more, compared to cloud BS So maybe i got 1h downtime more than i'd get on the cloud, but earning even $100k/hour is a preety decent pay
    Last edited by slawek22; 11-13-2012 at 08:47 AM.
      0 Not allowed!

  6. #131
    Join Date
    Nov 2000
    Location
    localhost
    Posts
    3,771
    In the 'theoretical worse case scenario' where every single machine is absolutely maxing out IO, then each client should still see 60-70MB/sec, yes, you are correct.
    This is the only number one should care about! Any serious startup, developer etc.. that are going to be scaling are going to be interested in this figure, what they might get is a bonus. Before developers would find a lower limit of their EBSs over time that werent in a degraded state and work from this, now they have another option with provision IOPS.

    You should put an SLA on that, i.e. we guarantee a floor of [some-IO-metric] over 99.9% of each calendar month, failure to meet this = next month free.
    MattF - Since the start..
      0 Not allowed!

  7. #132
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    Good point Matt, I'll raise that with the team here!
      0 Not allowed!

  8. #133
    Ran a test today using and using 6 nodes out of a batch we purchased a few days back.

    @ldn1 [~]# dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync
    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 53.4 s, 80.4 MB/s

    Edit: Ran a couple of minites later

    @ldn1 [~]# dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync

    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 74.7648 s, 57.4 MB/s

    Hmm...
    Last edited by DavidBee; 12-24-2012 at 06:52 PM.
      0 Not allowed!

  9. #134
    dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync
    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 8.68962 s, 494 MB/s

    There's one of mine but its not a Dediserve.

    I think one thing that cloud outfits may hopefully start doing in the future is actually telling us what the true contention for resources is.

    I'm gonna start some cloud services in a very small way soon and the first thing I am going to do is testing with the maximum number of clients I will put on any one single hypervisor and then publish figures based on the WORST case scenario that a client should accept.
    <<< Please see Forum Guidelines for signature setup. >>>
      0 Not allowed!

  10. #135
    Join Date
    Jul 2001
    Location
    Melbourne, AU
    Posts
    1,392
    Quote Originally Posted by DigitalDaz View Post
    dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync
    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 8.68962 s, 494 MB/s

    There's one of mine but its not a Dediserve.

    I think one thing that cloud outfits may hopefully start doing in the future is actually telling us what the true contention for resources is.

    I'm gonna start some cloud services in a very small way soon and the first thing I am going to do is testing with the maximum number of clients I will put on any one single hypervisor and then publish figures based on the WORST case scenario that a client should accept.
    I'm guessing your test isn't from a cloud VM and most likely from a standard VM.
    SERVSTRA | THE ENTERPRISE CLOUD SERVER & DEDICATED SERVER SPECIALISTS
    中国优先网络 - 最快到中国大陆
    14 world wide locations to choose from!
      0 Not allowed!

  11. #136
    Join Date
    Jul 2001
    Location
    Melbourne, AU
    Posts
    1,392
    Quote Originally Posted by David-Host Surf UK View Post
    Ran a test today using and using 6 nodes out of a batch we purchased a few days back.

    @ldn1 [~]# dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync
    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 53.4 s, 80.4 MB/s

    Edit: Ran a couple of minites later

    @ldn1 [~]# dd if=/dev/zero of=test bs=64k count=64k conv=fdatasync

    65536+0 records in
    65536+0 records out
    4294967296 bytes (4.3 GB) copied, 74.7648 s, 57.4 MB/s

    Hmm...
    And your point is? Are these results impacting on the running of your website? They look fine to me.
    SERVSTRA | THE ENTERPRISE CLOUD SERVER & DEDICATED SERVER SPECIALISTS
    中国优先网络 - 最快到中国大陆
    14 world wide locations to choose from!
      0 Not allowed!

  12. #137
    Quote Originally Posted by Servstra-Sales View Post
    I'm guessing your test isn't from a cloud VM and most likely from a standard VM.
    Its actually the first VM on a hypervisor that I will offer as a service but my point is that it is the only one on it at the moment. A true figure will be one that I give when I have 31 on it as thats the limit I will be putting on my HVs
    <<< Please see Forum Guidelines for signature setup. >>>
      0 Not allowed!

  13. #138
    Join Date
    Jul 2001
    Location
    Melbourne, AU
    Posts
    1,392
    Quote Originally Posted by DigitalDaz View Post
    Its actually the first VM on a hypervisor that I will offer as a service but my point is that it is the only one on it at the moment. A true figure will be one that I give when I have 31 on it as thats the limit I will be putting on my HVs
    Are you using a SAN or local storage?
    SERVSTRA | THE ENTERPRISE CLOUD SERVER & DEDICATED SERVER SPECIALISTS
    中国优先网络 - 最快到中国大陆
    14 world wide locations to choose from!
      0 Not allowed!

  14. #139
    And your point is? Are these results impacting on the running of your website? They look fine to me.
    Maybe David doesn't know all of the "advantages" of cheap virtualized enviroments yet, like unstable I/O.

    For us "mortals" it could be suprising when we discover that even I/O is broken by design in supposedly "superior" and "modern" cloud enviroments.
      0 Not allowed!

  15. #140
    Join Date
    Jan 2012
    Location
    UK
    Posts
    236
    there's a old saying which i will adapt to the "cloud" market.

    between HA, IO and price .... choose 2

    good IO + price .... local storage
    good HA + price ... shared SAN
    good IO and HA ... expensive dedicated SAN slice

    in theory Onapp Storage model tries to do all 3 at the same time, but time will prove how it will work.

    also please take in account i did not go into interconnect differences (1Gbe vs 10Gbe vs 4/8Gbe FC, etc)

    comparison between storage pricing in general (lets say SAS-7.2k)

    12 Tb of local storage = 1,500$ [ 0.125 $ per Mb ]
    12 Tb of non HA SAN on the cheap = 2,500$ [ 0.208 $ per Mb ]
    12 Tb of HA SAN on the cheap = 7500$ [ 0.625 $ per Mb ]
    12 Tb of full proper HA SAN = 20,000$ [ 1.666 $ per mb ]

    (these are rough numbers for comparison only and in no way represent real value)
    Last edited by hafthorr; 12-26-2012 at 11:54 PM.
      0 Not allowed!

  16. #141
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    For what it's worth (this being a thread about our solution ) We're moving all clouds to version 3.

    Good IO and HA - running NetApp clusters with all SSD trays with 10Gig multipath from each HV and dedicated aggregates per HV.
      0 Not allowed!

  17. #142
    Join Date
    Jan 2012
    Location
    UK
    Posts
    236
    Quote Originally Posted by dediserve View Post
    For what it's worth (this being a thread about our solution ) We're moving all clouds to version 3.

    Good IO and HA - running NetApp clusters with all SSD trays with 10Gig multipath from each HV and dedicated aggregates per HV.
    why do you get v3 and not me ?

    and did you mean 10Gig multipath from each Storage, and dedicated aggregates (guessing 2x1gbe) per HV ? or also 2x10Gbe on each HV ?
      0 Not allowed!

  18. #143
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    No 20gbps per hypervisor using two 10gig nics to the storage networks. 2gbps wouldn't be nearly enough.
      0 Not allowed!

  19. #144
    Join Date
    Jan 2012
    Location
    UK
    Posts
    236
    Quote Originally Posted by dediserve View Post
    No 20gbps per hypervisor using two 10gig nics to the storage networks. 2gbps wouldn't be nearly enough.
    yeah i noticed you use "fat" hypervisors.

    we actually opted to use smaller HV's and stick to 2Gbe / 4Gbe for now.

    trying to get a quote from mellanox about their new 10gbe switches for a few weeks now .......
      0 Not allowed!

  20. #145
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    I guess fat is relative 24 cores and 96GB ram isn't particularly fat these days
      0 Not allowed!

  21. #146
    Join Date
    Jan 2012
    Location
    UK
    Posts
    236
    Quote Originally Posted by dediserve View Post
    I guess fat is relative 24 cores and 96GB ram isn't particularly fat these days
    im waiting to test out the fattwin's with dual E5 / 128Gb setup, seems like a sensible configuration these days. and yes for that density 10Gbe becomes a real requirement.

    now i just wish HDS has some "entry" level SSD's in their price list
      0 Not allowed!

  22. #147
    Following on from DigitalDaz and slawek22, I'm also running into a performance issue that is apparently due to poor disk performance. For the record, I'm on their New York FusionIO platform which Dediserve put me on free of charge as an attempt to fix a problem and I remain very grateful for this. I'm particularly grateful for the time their support team has spent on me in the last month.

    As mentioned before by someone, I too am noticing that performance can be exceptional and yet can also be the complete opposite. On average every hour for a few minutes I’m experiencing incredibly poor performance. As an example, page loads of 25+ seconds that just involve executing 10 queries.

    I had Steve from Rack911 look into my server as his reputation is flawless here, so I take what he says very seriously.

    Steve is telling me that the performance issues I'm having with Dediserve are due to disk iowait. He discovered that 19% of my idle cpu is consumed by iowait and that it often spikes to 25%. He states that it is never acceptable to have iowait higher than your cpu usage. He also points out that a 200MB table took 5 minutes to repair when he would usually expect it to be under a minute.

    In response, Dediserve told me I should upgrade RAM/CPU but the fact is my 2 cores are near idle most of the time and I’m generally using less than half of my 1GB of ram. Basically, my traffic is seasonable and right now is the off-season.

    As the conversation continued with Dedi, I was told the IOwait was perfectly normal and that as I was on Fusion, it was proof positive that I can’t be seeing any IO issues.

    Does anyone have any thoughts on this? Am I missing something?
    Last edited by wimbly; 02-27-2013 at 07:35 PM.
      0 Not allowed!

  23. #148
    Join Date
    Apr 2009
    Location
    UK
    Posts
    824
    Quote Originally Posted by wimbly View Post
    Following on from DigitalDaz and slawek22, I'm also running into a performance issue that is apparently due to poor disk performance. For the record, I'm on their FusionIO platform which Dediserve put me on free of charge as an attempt to fix a problem and I remain very grateful for this. I'm particularly grateful for the time their support team has spent on me in general.

    As mentioned before by someone, I too am noticing that performance can be exceptional and yet can also be the complete opposite. On average every hour for a few minutes I’m experiencing incredibly poor performance. As an example, page loads of 25+ seconds that just involve executing 10 queries.

    I had Steve from Rack911 look into my server as his reputation is flawless here, so I take what he says very seriously.

    Steve is telling me that the performance issues I'm having with Dediserve are due to disk iowait. He discovered that 19% of my idle cpu is consumed by iowait and that it often spikes to 25%. He states that it is never acceptable to have iowait higher than your cpu usage. He also points out that a 200MB table took 5 minutes to repair when he would usually expect it to be under a minute.

    In response, Dediserve told me I should upgrade RAM/CPU but the fact is my 2 cores are near idle most of the time and I’m generally using less than half of my 1GB of ram. Basically, my traffic is seasonable and right now is the off-season.

    As the conversation continued with Dedi, I was told the IOwait was perfectly normal and that as I was on Fusion, it was proof positive that I can’t be seeing any IO issues.

    Does anyone have any thoughts on this? Am I missing something?
    Most clouds if not all have this kind of fluctuation with disk performance. With that money you could get a vps with ssd and your app will fly. Or even better 2 vps for redundancy.
      0 Not allowed!

  24. #149
    Quote Originally Posted by Losvre View Post
    Most clouds if not all have this kind of fluctuation with disk performance. With that money you could get a vps with ssd and your app will fly. Or even better 2 vps for redundancy.
    What is your opinion on iowait constantly being higher than cpu usage?

    BTW, this is on local storage, not SAN.
      0 Not allowed!

  25. #150
    Join Date
    Aug 2011
    Location
    Dub,Lon,Dal,Chi,NY,LA
    Posts
    1,839
    Hi Mark

    You never came back to us with additional data after being moved to fusion.

    Are you seeing any issues? We are here to help as always.

    As you might have noticed this thread is a few months old
      0 Not allowed!

Page 6 of 10 FirstFirst ... 3456789 ... LastLast

Similar Threads

  1. E5620 vs E5-2620 Performance / Benchmark comparison?
    By Leriss in forum Colocation, Data Centers, IP Space and Networks
    Replies: 16
    Last Post: 03-30-2012, 05:11 PM
  2. How to benchmark mysql io/performance
    By mixart in forum Dedicated Server
    Replies: 3
    Last Post: 02-12-2010, 06:42 PM
  3. Which CPU benchmark best approximates server performance?
    By Cripto in forum Dedicated Server
    Replies: 3
    Last Post: 02-10-2006, 07:24 PM
  4. running the benchmark test
    By aqi32 in forum Dedicated Server
    Replies: 11
    Last Post: 12-04-2004, 06:31 PM
  5. Scalable *nix based performance benchmark
    By Perlboy in forum Hosting Security and Technology
    Replies: 3
    Last Post: 10-21-2003, 03:56 AM

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •