3 ms·

LizardFS is Pretty Nice

2 points by Pneumaticat 8y ago

Annatar 8y ago
"However, it doesn’t write files whole – it breaks each into 64 MB chunks and distributes them across its drives." And what happens when the file is <= 64 MB? Does it run on illumos / SmartOS?
- Pneumaticat 8y ago
  According to the LizardFS whitepaper [1], the chunks can be up to 64 MB. So I believe if the file is smaller, the chunk is just as large as necessary. Looking at some chunk files on my system backs this up: root@rem> ls -l | head total 11702880 -rw-r----- 1 mfs mfs 33558528 Dec 12 18:17 chunk_ec_1_of_2_1_000000000000010C_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:12 chunk_ec_1_of_2_1_0000000000000179_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:11 chunk_ec_1_of_2_1_0000000000000187_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:26 chunk_ec_1_of_2_1_0000000000000194_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:06 chunk_ec_1_of_2_1_0000000000000197_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:36 chunk_ec_1_of_2_1_00000000000001A1_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:03 chunk_ec_1_of_2_1_00000000000001CE_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:13 chunk_ec_1_of_2_1_00000000000001D4_00000001.mfs -rw-r----- 1 mfs mfs 135168 Dec 12 18:05 chunk_ec_1_of_2_1_0000000000000217_00000001.mfs As for illumos/SmartOS -- maybe? The LizardFS cookbook [2] has a section for illumos, but there's nothing in it. It should probably work (it's C++), but there is probably going to have to be manual effort involved.
  - Annatar 8y ago
    Now do a test with a file size of exactly 1 non-NULL byte and report back. If the 1-byte file gets replicated across multiple disks, corrupt one of the chunks and report back.
    - syn0byte 8y ago
      Depends entirely on what replication goals you have setup. My simple cluster (not EC or XOR) mandates 2 copies per chunk at a minimum. My 1-byte file was unaffected by a single chunk corruption, even on the node in which the corruption originated.

2 points by Pneumaticat 8y ago

"However, it doesn’t write files whole – it breaks each into 64 MB chunks and distributes them across its drives." And what happens when the file is <= 64 MB? Does it run on illumos / SmartOS?

Pneumaticat 8y ago
According to the LizardFS whitepaper [1], the chunks can be up to 64 MB. So I believe if the file is smaller, the chunk is just as large as necessary. Looking at some chunk files on my system backs this up: root@rem> ls -l | head total 11702880 -rw-r----- 1 mfs mfs 33558528 Dec 12 18:17 chunk_ec_1_of_2_1_000000000000010C_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:12 chunk_ec_1_of_2_1_0000000000000179_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:11 chunk_ec_1_of_2_1_0000000000000187_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:26 chunk_ec_1_of_2_1_0000000000000194_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:06 chunk_ec_1_of_2_1_0000000000000197_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:36 chunk_ec_1_of_2_1_00000000000001A1_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:03 chunk_ec_1_of_2_1_00000000000001CE_00000001.mfs -rw-r----- 1 mfs mfs 69632 Dec 12 18:13 chunk_ec_1_of_2_1_00000000000001D4_00000001.mfs -rw-r----- 1 mfs mfs 135168 Dec 12 18:05 chunk_ec_1_of_2_1_0000000000000217_00000001.mfs As for illumos/SmartOS -- maybe? The LizardFS cookbook [2] has a section for illumos, but there's nothing in it. It should probably work (it's C++), but there is probably going to have to be manual effort involved.
- Annatar 8y ago
  Now do a test with a file size of exactly 1 non-NULL byte and report back. If the 1-byte file gets replicated across multiple disks, corrupt one of the chunks and report back.
  - syn0byte 8y ago
    Depends entirely on what replication goals you have setup. My simple cluster (not EC or XOR) mandates 2 copies per chunk at a minimum. My 1-byte file was unaffected by a single chunk corruption, even on the node in which the corruption originated.