From: 노홍찬 <fal...@cs...> - 2010-11-03 10:28:08
|
Dear all the developers, I recently tested several flash memory based SSDs (Intel x25-e, Intel x25- m, … ) by using IOmeter. In the same machine, I executed the most recent IOmeter to the same flash disk, using the same configuration as only random read requests on the basis of 4K I/O size. There was no big difference of IOPS performance between Linux (fedora 10) and Window (window server 8) when I set the number of outstanding I/O to 1. However, when I set the the number of outstanding I/O to a big number 32, then the IOPS number in window test result was almost 2 times of the IOPS number in linux test result. I used the same machine, the same storage device, and the same setting, but there was a big performance gap between the IOmeter test result in Window and in Linux. It was so strange to me, so I thought it’s because the asynchronous support of linux kernel might be much worse than that of window kernel. It will be very helpful, if you can give me an answer for this question. Why is the LINUX IOPS lower than the window IOPS in the same machine when the number of outstanding I/O is high? Best Regards Nate. |
Re: [Iometer-devel] A query for perfermance difference of
asynchronous I/O between Linux and Window.
From: Allen, W. <way...@in...> - 2010-11-03 16:16:05
|
Hi Nate, There have been some kernel changes to Linux over the past few years that has made IOMeter not behave correctly on that OS with > QD1. We expect to resolve this in the next release of IOMeter. In the mean time you can use a tool called FIO to perform your Linux-based tests. Best Regards, Wayne From: 노홍찬 [mailto:fal...@cs...] Sent: Wednesday, November 03, 2010 3:28 AM To: iom...@li... Subject: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Dear all the developers, I recently tested several flash memory based SSDs (Intel x25-e, Intel x25-m, … ) by using IOmeter. In the same machine, I executed the most recent IOmeter to the same flash disk, using the same configuration as only random read requests on the basis of 4K I/O size. There was no big difference of IOPS performance between Linux (fedora 10) and Window (window server 8) when I set the number of outstanding I/O to 1. However, when I set the the number of outstanding I/O to a big number 32, then the IOPS number in window test result was almost 2 times of the IOPS number in linux test result. I used the same machine, the same storage device, and the same setting, but there was a big performance gap between the IOmeter test result in Window and in Linux. It was so strange to me, so I thought it’s because the asynchronous support of linux kernel might be much worse than that of window kernel. It will be very helpful, if you can give me an answer for this question. Why is the LINUX IOPS lower than the window IOPS in the same machine when the number of outstanding I/O is high? Best Regards Nate. |
Re: [Iometer-devel] A query for perfermance difference of
asynchronous I/O between Linux and Window.
From: 노홍찬 <fal...@cs...> - 2010-11-04 03:04:42
|
Hi Wayne, Thank you very much for your kind reply. Can you specify the recent kernel problems not compatible with the current version of IOmeter in more detail, since in the first place I wanted to check the problems of asynchronous I/O in linux by using IOmeter benchmark. Is it the direct I/O or the native asynchronous I/O APIs? If the asynchronous I/O requests in FIO benchmark works fine, then I think I can use their method in programming an efficient asynchronous I/O module in Linux. Best Regards, Nate. From: Allen, Wayne [mailto:way...@in...] Sent: Thursday, November 04, 2010 1:16 AM To: 노홍찬; iom...@li... Subject: RE: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Hi Nate, There have been some kernel changes to Linux over the past few years that has made IOMeter not behave correctly on that OS with > QD1. We expect to resolve this in the next release of IOMeter. In the mean time you can use a tool called FIO to perform your Linux-based tests. Best Regards, Wayne From: 노홍찬 [mailto:fal...@cs...] Sent: Wednesday, November 03, 2010 3:28 AM To: iom...@li... Subject: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Dear all the developers, I recently tested several flash memory based SSDs (Intel x25-e, Intel x25- m, … ) by using IOmeter. In the same machine, I executed the most recent IOmeter to the same flash disk, using the same configuration as only random read requests on the basis of 4K I/O size. There was no big difference of IOPS performance between Linux (fedora 10) and Window (window server 8) when I set the number of outstanding I/O to 1. However, when I set the the number of outstanding I/O to a big number 32, then the IOPS number in window test result was almost 2 times of the IOPS number in linux test result. I used the same machine, the same storage device, and the same setting, but there was a big performance gap between the IOmeter test result in Window and in Linux. It was so strange to me, so I thought it’s because the asynchronous support of linux kernel might be much worse than that of window kernel. It will be very helpful, if you can give me an answer for this question. Why is the LINUX IOPS lower than the window IOPS in the same machine when the number of outstanding I/O is high? Best Regards Nate. |
Re: [Iometer-devel] A query for perfermance difference of
asynchronous I/O between Linux and Window.
From: Allen, W. <way...@in...> - 2010-11-04 05:05:49
|
I’m personally not familiar with the specifics on the kernel changes. FIO issues asynchronous IOs (as far as what I’ve seen it do). To achieve queue depth with FIO, you’ll need to set it up with the # of threads to match the queue depth you are looking for. Best Regards, Wayne From: 노홍찬 [mailto:fal...@cs...] Sent: Wednesday, November 03, 2010 8:04 PM To: Allen, Wayne; iom...@li... Subject: RE: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Hi Wayne, Thank you very much for your kind reply. Can you specify the recent kernel problems not compatible with the current version of IOmeter in more detail, since in the first place I wanted to check the problems of asynchronous I/O in linux by using IOmeter benchmark. Is it the direct I/O or the native asynchronous I/O APIs? If the asynchronous I/O requests in FIO benchmark works fine, then I think I can use their method in programming an efficient asynchronous I/O module in Linux. Best Regards, Nate. From: Allen, Wayne [mailto:way...@in...] Sent: Thursday, November 04, 2010 1:16 AM To: 노홍찬; iom...@li... Subject: RE: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Hi Nate, There have been some kernel changes to Linux over the past few years that has made IOMeter not behave correctly on that OS with > QD1. We expect to resolve this in the next release of IOMeter. In the mean time you can use a tool called FIO to perform your Linux-based tests. Best Regards, Wayne From: 노홍찬 [mailto:fal...@cs...] Sent: Wednesday, November 03, 2010 3:28 AM To: iom...@li... Subject: [Iometer-devel] A query for perfermance difference of asynchronous I/O between Linux and Window. Dear all the developers, I recently tested several flash memory based SSDs (Intel x25-e, Intel x25-m, … ) by using IOmeter. In the same machine, I executed the most recent IOmeter to the same flash disk, using the same configuration as only random read requests on the basis of 4K I/O size. There was no big difference of IOPS performance between Linux (fedora 10) and Window (window server 8) when I set the number of outstanding I/O to 1. However, when I set the the number of outstanding I/O to a big number 32, then the IOPS number in window test result was almost 2 times of the IOPS number in linux test result. I used the same machine, the same storage device, and the same setting, but there was a big performance gap between the IOmeter test result in Window and in Linux. It was so strange to me, so I thought it’s because the asynchronous support of linux kernel might be much worse than that of window kernel. It will be very helpful, if you can give me an answer for this question. Why is the LINUX IOPS lower than the window IOPS in the same machine when the number of outstanding I/O is high? Best Regards Nate. |