From: nikita <nikita>
Date: Tue, 20 Jun 2006 13:32:14 +0000 (+0000)
Subject: ll_commit_write(): add a comment (due to Andreas) explaining why Lustre wants writeba... 
X-Git-Tag: v1_8_0_110~486^2~1596
X-Git-Url: https://git.whamcloud.com/?a=commitdiff_plain;h=91ea7005b5d15575435dad7d3f39a4da0e970fee;p=fs%2Flustre-release.git

ll_commit_write(): add a comment (due to Andreas) explaining why Lustre wants writeback to be started earlier than "normal" file systems
---

diff --git a/lustre/llite/rw.c b/lustre/llite/rw.c
index 0bab99e..8c0a8d2 100644
--- a/lustre/llite/rw.c
+++ b/lustre/llite/rw.c
@@ -731,8 +731,26 @@ int ll_commit_write(struct file *file, struct page *page, unsigned from,
 
         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 
-        /* queue a write for some time in the future the first time we
-         * dirty the page */
+        /*
+         * queue a write for some time in the future the first time we
+         * dirty the page.
+         *
+         * This is different from what other file systems do: they usually
+         * just mark page (and some of its buffers) dirty and rely on
+         * balance_dirty_pages() to start a write-back. Lustre wants write-back
+         * to be started earlier for the following reasons:
+         *
+         *     (1) with a large number of clients we need to limit the amount
+         *     of cached data on the clients a lot;
+         *
+         *     (2) large compute jobs generally want compute-only then io-only
+         *     and the IO should complete as quickly as possible;
+         *
+         *     (3) IO is batched up to the RPC size and is async until the
+         *     client max cache is hit
+         *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
+         *
+         */
         if (!PageDirty(page)) {
                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
                                      LPROC_LL_DIRTY_MISSES);