Recently, Ive been researching using Linux at the center of an IP storage network. Specifically, I researched storage controllers (SATA) from vendors such as Broadcom, 3Ware and SuperMicro. When reviewing available performance data in the form of benchmarks , I normally tried to compare the controller’s performance on Linux against it’s performance on the Windows OS. In almost every benchmark I’ve read and digested, it seems that the controllers do not perform as well on Linux as they do in Windows. An example of this can be found in a white paper read through. That paper can be found here:
http://mgmt.datatag.org/sravot/paper/TCP_WAN_sr081604.pdf
The white paper is a research effort to overcome obstacles when performing data transfers over high speed WANS. It also deals with the last mile of the data transmission, which is to disk.
In one of the testing involving the SuperMicro SATA controller, it shows that the best write on Microsoft Windows is 85% BETTER than the best write on Linux. They attribute this obvious inconsistency to “better drivers and optimization in the Microsoft OS”. This is soooo frustrating. It seems that vendors are trying to attract Linux system builders and admins by developing drivers for their hardware on the Linux platform. This good! We want access to better and more modern hardware to run on our Linux powerhouses. Howerver, the performance on their hardware on Linux is no where near what one would achieve when they take the same card out of the system and stick it in an identical Windows box. Vendors, from my perspective, are not working close enough with the Linux kernel team to develop high quality drivers and software for their hardware. They appear to be just throwing stuff together and giving it to us and saying “Hey, guess what.. we got a driver for our XXX-455 Super SCSI Wonder Card”. Just go to this vendor and pay $1K US. But when you stick that card in your system, load the drivers and run bonnie++, you are very upset when you manage to get 200MB write per second across 12 drives JBOD when you know that you can get 400MB on Windows! I think the linux community needs to develop a standards body that addresses this issue and holds vendors accounting for delivering quality software and drivers for the Linux OS. There needs to be a Hardware quality Labs equivliant to the WHQL (Windows Hardware Quality Labs) that standardizes linux driver development and independantly tests and confirms the operating of hardware and their drivers on the Linux OS and makes vendors work to get a seal of approval. Does anyone have any thoughts?
Errol