Hi Yehuda, Thanks for your answers. The situation behind these questions is we found that an application data load is > 512KB and <4MB (maybe < 1 MB more precisely). The GET request usually requires 2 rados read ops. If there is a way we can configure the RGW_MAX_CHUNK_SIZE to a number between 512KB and 1MB, only 1 read op will be needed. So we can get total latency improved much even the first byte latency increased very little. Thanks, Ray On 2/21/14, 4:46 AM, "Yehuda Sadeh" <yehuda@xxxxxxxxxxx> wrote: >Apparently I missed including the mailing list in the response. > >---------- Forwarded message ---------- >From: Yehuda Sadeh <yehuda@xxxxxxxxxxx> >Date: Thu, Feb 20, 2014 at 9:04 AM >Subject: Re: [rgw] increase the first chunk size >To: Ray Lv <raylv@xxxxxxxxxxxxx> > > >On Thu, Feb 20, 2014 at 5:08 AM, Ray Lv <raylv@xxxxxxxxxxxxx> wrote: >> Hi, >> >> Currently, the first chunk size of a radosgw object is 512KB. >> >> Here is a case that most of data workload gets to radosgw is ~ 4MB. >>With the >> current first chunk size, each radosgw object is stripped to two chunks >> (512K + 3.5M). And we're using several large disks on each host with >>40TB >> capacity. So there will be 10 millions of files on each host. If the >>first >> chunk size is increased to 4MB, the number of files on each host will be >> reduced by 50%. It will be benifitial to performance of read because of >> reduced dcache and inode cache footprint in main memory (in other words, >> increased cache hit ratio). >> >> The questions are: >> >> What's the rationale behind for the current first chunk size? > >The head side conforms to the read chunk side. The gateway reads in >512k chunks. This is the basic read unit, and when accessing an object >we access the head only once and read the entire data + all its >attributes in one compound rados operation. This is done to ensure >atomicity. We don't have any knowledge as to whether the operation is >deemed to fail later on (e.g., insufficient permissions), so it's >going to read it anyway. > >> Are there any side effects if it is increased to 4MB? > >Will require reading the head in a single operation, which can be a >problem with regard to concurrency, add latency to all operations >(will take more time to stream data back to the client). Unauthorized >requests will use more resources. > > >Yehuda _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |