November 2, 2010 at 14:33Comparison of dedicated server processor speeds
(on a graph along the abscissa axis a logarithmic scale)
During the discussion of the announcement of rented servers based on atoms of 1500r each, the question arose of how much the atom is slower than other platforms.
The question was interesting. Here are the test results. As a test, sysbench was used (it is in many Linux distributions, including Ubuntu and Squeeze).
The tests distinguish two cases - single-threaded and multi-threaded load. A typical load on a visited web server is multithreaded. A typical load for a single single-threaded application (for example, gzip on big data) is single-threaded.
It is clearly seen that atoms significantly lose to Core / Xeon processors, which turn out to be one and a half to two times faster in one thread than an atom with two cores and hypertreading.
Another interesting observation - on the atom, the 32-bit and 64-bit modes show the same thing; on all other processors, the 64-bit architecture outperforms the 32-bit one.
The reference was called nice -20 sysbench --test = cpu --cpu-max-prime = 40000 --num-threads = X run (X is the number of threads, from 1 to 16).
On the chart, this data. Please note the semi-logarithmic scale (in the linear difference between the atoms and the rest was so great that it made no sense to compare the other processors with each other). It is also seen that CoreDuo is the fastest in single-threaded execution (which is understandable, because it is a desktop processor, sharpened by resource-intensive single-threaded applications, like games). Xeons show their speed with maximum parallelization of tasks (typical server load).
Of particular interest is the growth that HT provides (with HT there are two models - Atom D150 and Xeon 5620):
This comparison applies only to processors and does not fully characterize the platform. Atoms come with 2GB of memory, peels - 6GB, zeons - 12 and 24GB, respectively; also varies equipment discs.
PS Since the logarithmic scale raised a lot of questions, here's a different view of the same data: processor performance is relatively better (i.e. 100%). The scale is linear.
During the discussion of the announcement of rented servers based on atoms of 1500r each, the question arose of how much the atom is slower than other platforms.
The question was interesting. Here are the test results. As a test, sysbench was used (it is in many Linux distributions, including Ubuntu and Squeeze).
The tests distinguish two cases - single-threaded and multi-threaded load. A typical load on a visited web server is multithreaded. A typical load for a single single-threaded application (for example, gzip on big data) is single-threaded.
It is clearly seen that atoms significantly lose to Core / Xeon processors, which turn out to be one and a half to two times faster in one thread than an atom with two cores and hypertreading.
Another interesting observation - on the atom, the 32-bit and 64-bit modes show the same thing; on all other processors, the 64-bit architecture outperforms the 32-bit one.
The reference was called nice -20 sysbench --test = cpu --cpu-max-prime = 40000 --num-threads = X run (X is the number of threads, from 1 to 16).
| Model | Atom D510 1.66 GHz | Core2Duo E8400 3.0GHz | Core2Quad Q8300 2.5GHz | 2 x Xeon 5504 2.0 GHz | 2 x Xeon 5620 |
| Cores | 2 + HT | 2 | 4 | 8 | 8 + HT |
| Single-Threaded Computing (32/64) | 412.83 / 412.82 | 61.89 / 54.98 | 73.47 / 66.03 | 90.31 / 83.07 | 75.28 / 76.43 |
| Multithreaded Computing (32/64) | 124.55 / 124.57 | 30.91 / 27.49 | 18.43 / 16.56 | 11.29 / 10.45 | 8.91 / 7.21 |
On the chart, this data. Please note the semi-logarithmic scale (in the linear difference between the atoms and the rest was so great that it made no sense to compare the other processors with each other). It is also seen that CoreDuo is the fastest in single-threaded execution (which is understandable, because it is a desktop processor, sharpened by resource-intensive single-threaded applications, like games). Xeons show their speed with maximum parallelization of tasks (typical server load).
Of particular interest is the growth that HT provides (with HT there are two models - Atom D150 and Xeon 5620):
| CPU | Atom D510 1.66 GHz | 2 x Intel Xeon 5620 |
| without HT (multithreaded) | 206.47 | 11.35 |
| with HT (multithreaded) | 124.57 | 7.21 |
| Speed increase (%) | 65.7% | 57.4% |
This comparison applies only to processors and does not fully characterize the platform. Atoms come with 2GB of memory, peels - 6GB, zeons - 12 and 24GB, respectively; also varies equipment discs.
PS Since the logarithmic scale raised a lot of questions, here's a different view of the same data: processor performance is relatively better (i.e. 100%). The scale is linear.