Why haven't NetMap or MegaPipe been merged?
Why haven't NetMap or MegaPipe been merged? I am wondering because they seem to be big wins for performance, but neither has been upstreamed.
On Mon, 12 Feb 2018 12:48:11 -0500, Demi Obenour said:
Why haven't NetMap or MegaPipe been merged? I am wondering because they seem to be big wins for performance, but neither has been upstreamed.
It helps if you help identify things for those who haven't been deeply involved. Are these IP network things, or fiberchannel stuff, or Infiniband, or something else? Where is *mergeable* source for them? (usually, the "why isn't this already merged?" is "because the code is a steaming pile of dingo's kidneys"). Are there actual measurements rather than just "seem to be big wins"? What downsides are there? What configurations does this become a big performance *loss*? Are the downsides documented? And that's just the start of questions that will get asked before it gets merged....
On Mon, Feb 12, 2018 at 12:48:11PM -0500, Demi Obenour wrote:
Why haven't NetMap or MegaPipe been merged? I am wondering because they seem to be big wins for performance, but neither has been upstreamed.
Has anyone actually submitted the patches upstream? That's usually the first step :)
Not sure. The NetMap team at least told me that upstream didn't want their code. This might be because NetMap requires special vendor-provided (but still open source, if I understand correctly) drivers for optimum performance, or it might be because they use a single codebase to support Windows, Linux, and FreeBSD, if I understand correctly. Hide quoted text On Feb 12, 2018 2:21 PM, "Greg KH" <greg@kroah.com> wrote:
On Mon, Feb 12, 2018 at 12:48:11PM -0500, Demi Obenour wrote:
Why haven't NetMap or MegaPipe been merged? I am wondering because they seem to be big wins for performance, but neither has been upstreamed.
Has anyone actually submitted the patches upstream? That's usually the first step :)
On Tue, Feb 13, 2018 at 01:00:39PM -0500, Demi Obenour wrote:
Not sure. The NetMap team at least told me that upstream didn't want their code. This might be because NetMap requires special vendor-provided (but still open source, if I understand correctly) drivers for optimum performance, or it might be because they use a single codebase to support Windows, Linux, and FreeBSD, if I understand correctly.
Well, "special" drivers, as well as a compatibility layer for other operating systems are a very good reason to not get the code merged into the kernel. No subsystem would ever accept such an odd requirement. greg k-h
participants (3)
-
Demi Obenour -
Greg KH -
valdis.kletnieks＠vt.edu