Jschiller/fix minfs pants #142
Conversation
experiment with fixed-size buffer pool using a channel
Rework the fuse library to have less overhead. Make it a new library with clear lineage.
#5) * Removed type switching from serve.go and replaced it with a Servlet handler passed to Conn.Read(). * For now, explicitly disabling support for forget requests. We'll revisit this later.
#6) * Selectively clear allocator bytes since most of them are immediately overridden. Added ThreadUnsafe mount option to control locking (not used yet). Removed all locking, it doesn't appear necessary on our supported platforms, second opinion needed though. Added some additional ops to the unsupported list. * Added comments about allocator zeroing and removed an unused field.
…ome legitimate user fs ops to fail.
|
What happened here?-) |
|
User error happened - github defaults to your repo every time I create a pull request :/ |
|
I saw some of those changes. We should have a chat. I feel you're fixing little nits and missing the bigger things like #35 -- with |
|
The only important change in all of this is reducing the amount of allocation, and you're right everything else is basically nits. We're focusing on other code right now, after which we plan to scope out if we'll use fuse or our own vfs. Vmsplice is of interest though and I'll DM you if/when we decide to make more perf changes. Are you hoping for a PR or do you have this work planned out already? |
|
Well, I have a rough idea of how to use With vmsplice, there will be practically no "buffer management" allocations anymore, just a pool of large regions with multiple messages packed in each, to be reused after a refcount drops to zero. Oh and I like what you did with avoiding the type switch, I might migrate to that too. |
No description provided.