On Thu, May 09, 2013 at 02:56:50PM +0800, Fam Zheng wrote:
> On Thu, 05/09 08:41, Stefan Hajnoczi wrote:
> > On Fri, May 03, 2013 at 04:00:09PM +0800, Fam Zheng wrote:
> > > + cache = curl_find_cache(s, aio_base, aio_bytes);
> > > + if (cache) {
> > > + curl_complete_io(s, acb, cache);
> > > + return;
> > > }
> >
> > What is the point of the cache? Can you split it into a separate patch?
>
> The cache is for prefetch. Data is fetched by 256k chunks using libcurl
> and stored in cache to fill future io request, reducing overall http
> request overhead.
That sounds like a performance optimization which is not part of fixing
broken block/curl.c :).
But why bother doing this - the guest kernel uses read-ahead anyway to
fetch more data than the userspace read(2) requested? This cache
amplifies the read-ahead and potentially wastes (e.g. random-read
workload). Let the guest decide how much read-ahead is appropriate.
Stefan
