Uncategorized

Comparing Intel Dual-Core and Quad-Core Using VMmark

I have been running VMmark on a wide variety of systems in preparation for the benchmark’s imminent release. During this exercise, I was able to measure the performance difference between Intel’s dual-core (Woodcrest) and quad-core (Clovertown) processors using otherwise identical HP Proliant DL380G5 systems. One system contained two dual-core Intel Xeon 5150 processors (four cores total) running at 2.66 GHz. The other system contained two quad-core Intel Xeon 5355 processors (eight cores total) running at 2.66 GHz. Each system contained 32GB of memory and was connected to the same VM library hosted on an EMC CX-300 storage array. Both systems were running ESX Server 3.0.1.

The figure below shows the benchmark scores as the number of tiles was increased until CPU saturation was reached for each system. The dual-core system became fully saturated when running three tiles whereas the quad-core system became saturated at five tiles.

Quadblog2

The systems achieved nearly identical benchmark scores for one and two tiles. This behavior was expected since neither system had exhausted its CPU resources. At three tiles, the dual core system was fully utilized, which limited the score, while the quad core systems continued to scale well due to the additional available CPU resources. The quad-core system delivered a 28% higher score with three tiles. The quad-core system became saturated at five tiles and ultimately achieved a score 70% higher than the dual-core system while supporting 67% more VMs.

Many people have asked how much additional performance quad-core processors can provide over their dual-core cousins. These results should help to answer that question while demonstrating the value of having a representative benchmark for mixed virtualization workloads.

Comments

28 comments have been added so far

  1. VMware: vergelijking ESX 3.0.1 op dual- en quad-core CPUs

    VMware heeft een benchmark laten zien van hun VMmark utility. Hiermee kan de belasting van een ESX host getest worden en kunnen verschillende configuraties vergeleken worden. Tijdens het testen van deze nieuwe/vernieuwde tool, hebben ze Intels d…

  2. Speaking in terms of performance (not in security) what do you think of a VI3 environment formed by two IBM System x3950 32-way ? A environment formed by 08 08-way machines would be better in terms of performance ?

  3. Bruce,
    this is interesting but there is no mention about memory setup. Usually in a real-life scenario supporting 67% more vm’s would require a substantial amount of additional memory in the system and most servers has limited memory slots available (especially for VI3 configurations). So my comment is that while from a pure engineering perspective it is interesting to notice how a given workload scales across the cpu subsystem…. customers should be looking at a more holistic view which will determine the real price/performance of the platform.
    Massimo.

  4. Um…I believe he mentions the config of each server had 32GB of memory. Were you asking for a more detailed breakdown of what type of memory? How memory was allocated?

  5. The main problem with this configuration is that the machine is far more expensive with 32GB of RAM than with 16GB, as the densest DIMMs are required.
    1 DL380 with 32GB of RAM costs about 2 DL380 with 16GB of RAM.
    It would be good to have something like $/performance metric (like $/TPC).

  6. Ronaldo,
    You make an important point by bringing up pricing. We considered having a price metric included within VMmark, but ultimately decided against it for several reasons. First, pricing metrics are hard to define and verify. Ultimately, pricing metrics can cause vendors to use configurations in the benchmarks that are difficult for customers to easily evaluate. One example from TPC benchmarks is the pairing of expensive, high-end servers with cheap low-end storage configured in RAID 0, something no sane user would do in production. Also, system discounts tend to vary from customer to customer and are unlikely to match those given in a benchmark configuration.
    We felt that by providing a reliable performance metric with VMmark, we would enable customers to make more informed decisions. One should evaluate not just performance but also factors including price, power consumption, datacenter space, and system redundancy for HA. VMmark addresses the longstanding lack of perfomance data. Letting each customer price systems according to their needs should provide the most benefit.

  7. Claiton,
    VMmark was design to answer these types of questions. Unfortunately, I don’t have the types of systems you mention in my lab, so I can’t give you a definitive answer. I would suggest that you ask your IBM rep.

  8. My staff send me the following vmware blog related to dual-core and quad-core comparison using VMmark but I do not think that it is a fair comparison since it uses a server with 2 x quad-core versus a server with 2 x dual-core. What I think is a fair comparison will be 1 x quad-core versus 2 x dual-core and 2 x quad-core versus 4 x dual-core. Will we consider running the VMmark with those two configurations that I am proposing?

  9. I am also interested in the comparison between 2 X Dual and 1 X Quad CPU’s I am ordering some new servers and the price is about the same for 1 – E5255 Quad cpu as 2 – Dual core 5160 cpu’s.
    I would like to know which to buy? Would a quad cpu be faster?

  10. I am also interested in the comparison between 2 X Dual and 1 X Quad CPU’s I am ordering some new servers and the price is about the same for 1 – E5255 Quad cpu as 2 – Dual core 5160 cpu’s.
    I would like to know which to buy? Would a quad cpu be faster?

  11. I agree that a comparison between a single quad-core CPU and two dual-core CPUs would be interesting. I haven’t done that test myself. By releasing VMmark and working with our partners, we hope to enable them to generate results like this. I’d suggest you ask your sales rep for results on the particular systems you are interested in.
    Also, I don’t understand why one would consider the comparison described above as unfair. It merely provides performance data for two different CPU configurations of the same server. One can use this information to properly configure a system according to their needs. To illustrate, I went to the HP web site and priced DL380G5 systems similar to what I tested. The quad-core system costs only about 4% more than the dual core system ($18911 for the dual-core vs. $19651 quad-core) but provides 70% better throughput. Ronaldo pointed out that two dual-core systems configured with 16GB each might be a better value. According to the HP web site that system costs $12311 with 2GB DIMMs or $14113 with 4GB DIMS. Buying two of either of those systems would cost more than a single quad-core system with 32GB. The price/performace difference is close enough that one should probably consider other factors such as power, space, and cooling costs or the need for system redundancy for HA purposes.
    The wonderful thing about VMmark is that we now have a benchmark with which to generate the data necessary to evaualate these types of tradeoffs. It should be fun as more and more data becomes available.

  12. I’m sorry but i don’t understand what is tiles (French user inside :))?
    If there someone to explain me what it is?
    Thanks

  13. Can we do a comparison of a PE 2950 with 2 quads against a PE 6850 with 4 duals? I appreciate people being interested in the 1 quad vs 2 duals but I think most of us buying big boxes to load VM on are getting the 4 socket boxes (at least I am). Is it the best choice?

  14. just a note anyone trying to look at the dell vmark pdf. Remove the last period and you will be able to access the file.

  15. Would you please answer me shortly whether i should buy the Intel Quad core Q6600/2.4GHz/8MB or the Dual core E-6850/3.0GHz/4MB ?
    Thanks,

  16. Yaron,
    We don’t have the equipment to run exhaustive tests across processor and system families. However, we have been working closely with our partners to generate a wide range of benchmark results. I suggest that you check our newly unveiled VMmark results page at http://www.vmware.com/products/vmmark/results.html. You may find the data you need (or a reasonable substitute). If not, please ask your OEM for the data.

  17. With Regards to Quad vs Dual.
    IT scales lineraly, but it isnt exactly. Meaning that 2x 2.66 dual cores (4 cores), actually preform better thatn 1x quad 2.66. But not at a substantial level. the goal with the Vmwares are to have as many cpu’s as possible. Going from 16 to 32GB of ram, cost about 3000 on a new system config. YET 32GB of ram, adds easily 16x1GB servers, with very little tweaking. or 4x4GB ram servers with little tweaking. NOW with a dualx2core rig, you are going to have much more cpu contention then with a rig with dualx4core cpu’s. Cores in this matter actually make a difference. and price wise, not even close, considering that Dell for instance, prices the quads within 10% of their Duals, and mostly Cheaper tahn their duals..

  18. And to Kyle.. I bought a Quad core 6600.. put on my esx server. and WOW.. IT makes a huge difference (home lab). The cores help.. you notice immediatly after 4VM’s are running. Or if you have VM’s that are SMP. in production we noticed .. we tried 2x 5160’s 2xdual core. running Citrix. the servers were heavy use servers. had to use SMP. now with 2 ctx servers, it was ok, the moment we added 1 more server. BOOM, immediate impact. We upgraded the cpu’s to 2x quad’s 2.66 GHZ. and i tell you. we are running 4Ctx servers on this server, YET only at 55-60 % cpu utilizatoin. (to give idea, 1 Physical ctx server, gets SAPPED at 22 uers, versus on VMware we are getting 24-26 usersx 4 servers on SAME physical class server (w\ the 2x quads, and 16gb ram).
    So that alone proved to me the scalability of the Quads..
    Some ESX server installs are consilidation of older/legacy servers. Some like the one we use, is to push the hardware limitations, and get more per U

  19. Just echoing some of the comments I’ve read here…. I’d love to see this test run again with the original processor configuration (Dual Quad) but with 16GB of RAM instead of 32GB. The cost difference between eight 2GB DIMMS and eight 4GB DIMMS (for the HP DL360 G5) is quite significant!

  20. We do not have the equipment to carry out extensive testing across the processor and system families. However, we have been working closely with our partners to generate a wide range of benchmark results. I suggest you see our newly unveiled VMmark results page at http://www.vmware.com / products / VMmark / results.html. You can find the data you need (or a reasonable substitute). If not, contact your OEM for the data.

  21. Now, there are cheaper ways to move this many transactions on a virtual platform, but we weren’t building a cargo-hauling truck, we were building a race car to prove a point – scaling out with virtual machine building blocks is the best way to overcome multi-core scaling limitations of current applications.

Leave a Reply

Your email address will not be published. Required fields are marked *