Hi, ![]() I
am running TM using lazy (LL) on the the stamp benchmarks. The threads
keep getting stuck at commit because one thread acquires the write
token, begins flushing its buffer, and then stops without finishing the
buffer flush and never releases the token. This doesn't happen
immediately (many commits occur without a problem) but it's happening
consistently in intruder, kmeans, genome, and ssca2. I have
XACT_MEMORY=true, XACT_EAGER_CD=false, and XACT_LAZY_VM=true, all set at
compile time. The benchmarks run fine under eager and we are avoiding
malloc/other syscalls in transaction. Has anyone seen this before? Or
have any ideas what might be happening?
Thanks in advance,
Gwen
|