Re: [Qemu-devel] [PATCH v2] XBRLE page delta compression for live migration of large memory apps

[Stefan Hajnoczi]

On Wed, Jul 06, 2011 at 02:01:58PM +0200, Shribman, Aidan wrote:
> XBRLE does not suite all scenarios but has been proven highly beneficial in
> for VMs running payloads such as: SAP ERP systems; VLC transcoding; LMbench
> memory write benchmarks.

Another way of looking at this patch is as a profiling tool for identifying
workloads with poor TLB and cache behavior.  The workloads that benefit from
this patch are dirtying memory in a sparse fashion, touching many pages but
making only small changes.  These workloads are using data structures and
algorithms that are simply not TLB/cache-efficient.

Instead of spending effort in live migration compensating for this poor
workload behavior, why not fix the workload?  The benefits are much greater:
poor TLB/cache usage affects the workload all the time, not just during live
migration and not just under virtualization.  By fixing the workload you will
also get faster live migration.

That said, if we can improve live migration then this is a good thing.  The
challenge is that this optimization is speculative.  You need to do a lot of
work up-front: copying all pages through the cache and hoping their xor/rle
representation will be <1/3 TARGET_PAGE_SIZE.  If there is a cache miss or the
xor/rle representation is not small enough then it's back to square one.

Any thoughts on reducing the overhead and making xbrle on by default?

> Work is based on research results published VEE 2011: Evaluation of Delta
> Compression Techniques for Efficient Live Migration of Large Virtual Machines
> by Benoit, Svard, Tordsson and Elmroth.

I will read your paper.  Did you try unconditionally applying a cheap
compression algorithm like the one Google recently published?  That way you
just compress everything and don't need to keep the cache around:

http://code.google.com/p/snappy/
http://www.hypertable.org/doxygen/bmz_8h.html

> +static int save_xbrle_page(QEMUFile *f, uint8_t *current_data,
> +        ram_addr_t current_addr, RAMBlock *block, ram_addr_t offset, int 
> cont)
> +{
> +    int cache_location = -1, slot = -1, encoded_len = 0, bytes_sent = 0;
> +    XBRLEHeader hdr = {0};
> +    CacheItem *it;
> +    uint8_t *xor_buf = NULL, *xbrle_buf = NULL;
> +
> +    /* get location */
> +    slot = cache_is_cached(current_addr);
> +    if (slot == -1) {
> +        acct_info.xbrle_cache_miss++;
> +        goto done;
> +    }
> +    cache_location = cache_get_cache_pos(current_addr);
> +
> +    /* abort if page changed too much */
> +    it = cache_item_get(cache_location, slot);
> +
> +    /* XOR encoding */
> +    xor_buf = (uint8_t *) qemu_mallocz(TARGET_PAGE_SIZE);

Zeroing unnecessary here.

> +    xor_encode(xor_buf, it->it_data, current_data);
> +
> +    /* XBRLE (XOR+RLE) encoding (if we can ensure a 1/3 ratio) */
> +    xbrle_buf = (uint8_t *) qemu_mallocz(TARGET_PAGE_SIZE);

Why TARGET_PAGE_SIZE when the actual size is TARGET_PAGE_SIZE/3?

Zeroing unnecessary here.

> +    encoded_len = rle_encode(xor_buf, TARGET_PAGE_SIZE, xbrle_buf,
> +            TARGET_PAGE_SIZE/3);
> +
> +    if (encoded_len < 0) {
> +        DPRINTF("XBRLE encoding oeverflow - sending uncompressed\n");

s/oeverflow/overflow/

> +        acct_info.xbrle_overflow++;
> +        goto done;
> +    }
> +
> +    hdr.xh_len = encoded_len;
> +    hdr.xh_flags |= ENCODING_FLAG_XBRLE;
> +
> +    /* Send XBRLE compressed page */
> +    save_block_hdr(f, block, offset, cont, RAM_SAVE_FLAG_XBRLE);
> +    qemu_put_buffer(f, (uint8_t *) &hdr, sizeof(hdr));
> +    qemu_put_buffer(f, xbrle_buf, encoded_len);
> +    acct_info.xbrle_pages++;
> +    bytes_sent = encoded_len + sizeof(hdr);
> +    acct_info.xbrle_bytes += bytes_sent;
> +
> +done:
> +    qemu_free(xor_buf);
> +    qemu_free(xbrle_buf);
> +    return bytes_sent;
> +}
> 
>  static int is_dup_page(uint8_t *page, uint8_t ch)
>  {
> @@ -107,7 +486,7 @@ static int is_dup_page(uint8_t *page, uint8_t ch)
>  static RAMBlock *last_block;
>  static ram_addr_t last_offset;
> 
> -static int ram_save_block(QEMUFile *f)
> +static int ram_save_block(QEMUFile *f, int stage)
>  {
>      RAMBlock *block = last_block;
>      ram_addr_t offset = last_offset;
> @@ -128,28 +507,27 @@ static int ram_save_block(QEMUFile *f)
>                                              current_addr + TARGET_PAGE_SIZE,
>                                              MIGRATION_DIRTY_FLAG);
> 
> -            p = block->host + offset;
> +            p = qemu_mallocz(TARGET_PAGE_SIZE);

Where is p freed when use_xbrle is off?

You should not introduce overhead in the case where use_xbrle is off.  Please
make sure the malloc/memcpy only happens if the page is added to the cache.

> +static int load_xbrle(QEMUFile *f, ram_addr_t addr, void *host)
> +{
> +    int ret, rc = -1;
> +    uint8_t *prev_page, *xor_buf, *xbrle_buf;
> +    XBRLEHeader hdr = {0};
> +
> +    /* extract RLE header */
> +    qemu_get_buffer(f, (uint8_t *) &hdr, sizeof(hdr));
> +    if (!(hdr.xh_flags & ENCODING_FLAG_XBRLE)) {
> +        fprintf(stderr, "Failed to load XBRLE page - wrong compression!\n");
> +        goto done;
> +    }
> +
> +    if (hdr.xh_len > TARGET_PAGE_SIZE) {
> +        fprintf(stderr, "Failed to load XBRLE page - len overflow!\n");
> +        goto done;
> +    }
> +
> +    /* load data and decode */
> +    xbrle_buf = (uint8_t *) qemu_mallocz(TARGET_PAGE_SIZE);
> +    qemu_get_buffer(f, xbrle_buf, hdr.xh_len);

Why allocate TARGET_PAGE_SIZE instead of hdr.xh_len and why zero it when
qemu_get_buffer() will overwrite it?

> +
> +    /* decode RLE */
> +    xor_buf = (uint8_t *) qemu_mallocz(TARGET_PAGE_SIZE);

Again there is no need to zero the buffer.

Stefan